Applied Nonlinear Control

Nguyen Tan Tien - 2002.3

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

C.1 Introduction
Why Nonlinear Control ? Nonlinear control is a mature subject with a variety of powerful methods and a long history of successful industrial applications ⇒ Why so many researchers have recently showed an active interest in the development and applications of nonlinear control methodologies ?

Improvement of existing control systems Linear control methods rely on the key assumption of small range operation for the linear model to be valid. When the required operation range is large, a linear controller is likely to perform very poorly or to be unstable, because the nonlinearities in the system cannot be properly compensated for. Nonlinear controllers may handle the nonlinearities in large range operation directly. Ex: pendulum

Analysis of hard nonlinearities One of the assumptions of linear control is that the system model is indeed linearizable. However, in control systems, there are many nonlinearities whose discontinuous nature does not allow linear approximation. Ex: Coulomb friction, backlash

Dealing with model uncertainties In designing linear controllers, it is usually necessary to assume that the parameters of the system model are reasonably well known. However in many control problems involve uncertainties in the model parameters. Nonlinearities can be intentionally introduced into the control part of a control system so that model uncertainties can be tolerated. Two classes of nonlinear controllers for this purpose are robust controllers and adaptive controllers. Ex: parameter variations

Design simplicity Good nonlinear control designs may be simpler and more intuitive than their linear counterparts. Ex: & x = Ax + Bu & x = f + gu

___________________________________________________________________________________________________________ 1 Chapter 1 Introduction

Applied Nonlinear Control

Nguyen Tan Tien - 2002.3

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

2. Phase Plane Analysis
Phase plane analysis is a graphical method for studying second-order systems. This chapter’s objective is to gain familiarity of the nonlinear systems through the simple graphical method. 2.1 Concepts of Phase Plane Analysis 2.1.1 Phase portraits The phase plane method is concerned with the graphical study of second-order autonomous systems described by & x1 = f1 ( x1 , x 2 ) & 2 = f 2 ( x1 , x 2 ) x where x1 , x 2 : states of the system f1 , f 2 : nonlinear functions of the states Geometrically, the state space of this system is a plane having x1 , x 2 as coordinates. This plane is called phase plane. The solution of (2.1) with time varies from zero to infinity can be represented as a curve in the phase plane. Such a curve is called a phase plane trajectory. A family of phase plane trajectories is called a phase portrait of a system. Example 2.1 Phase portrait of a mass-spring system_______
& x

The nature of the system response corresponding to various initial conditions is directly displayed on the phase plane. In the above example, we can easily see that the system trajectories neither converge to the origin nor diverge to infinity. They simply circle around the origin, indicating the marginal nature of the system’s stability. A major class of second-order systems can be described by the differential equations of the form && = f ( x, x) & x (2.3)

(2.1a) (2.1b)

In state space form, this dynamics can be represented & with x1 = x and x 2 = x as follows & x1 = x2 & x2 = f ( x1 , x 2 ) 2.1.2 Singular points A singular point is an equilibrium point in the phase plane. Since an equilibrium point is defined as a point where the & system states can stay forever, this implies that x = 0 , and using (2.1)  f1 ( x1 , x 2 ) = 0   f 2 ( x1 , x 2 ) = 0 (2.4)

For a linear system, there is usually only one singular point although in some cases there can be a set of singular points.
x

k =1
0

Example 2.2 A nonlinear second-order system____________
& x
9

m =1
(a ) (b)
-6 -3

6
convergence 3 area

Fig. 2.1 A mass-spring system and its phase portrait The governing equation of the mass-spring system in Fig. 2.1 is the familiar linear second-order differential equation

3

6

x

unstable

&& + x = 0 x

(2.2)

-3

Assume that the mass is initially at rest, at length x0 . Then the solution of this equation is x(t ) = x0 cos(t ) & x(t ) = − x0 sin(t ) Eliminating time t from the above equations, we obtain the equation of the trajectories
2 & x + x 2 = x0 2

-6
to infinity

divergence area

-9

Fig. 2.2 A mass-spring system and its phase portrait

& Consider the system && + 0.6 x + 3x + x 2 = 0 whose phase x portrait is plot in Fig. 2.2.
The system has two singular points, one at (0,0) and the other at (−3,0) . The motion patterns of the system trajectories in the vicinity of the two singular points have different natures. The trajectories move towards the point x = 0 while moving away from the point x = −3 .
__________________________________________________________________________________________

This represents a circle in the phase plane. Its plot is given in Fig. 2.1.b.
__________________________________________________________________________________________

___________________________________________________________________________________________________________ 1 Chapter 2 Phase Plane Analysis

Applied Nonlinear Control

Nguyen Tan Tien - 2002.3

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

Why an equilibrium point of a second order system is called a singular point ? Let us examine the slope of the phase portrait. The slope of the phase trajectory passing through a point ( x1 , x 2 ) is determined by dx2 f (x , x ) = 2 1 2 dx1 f1 ( x1 , x 2 ) (2.5)

f ( x1 , x2 ) = f ( x1 ,− x 2 ) f ( x1 , x2 ) = − f ( x1 ,− x2 )

⇒ symmetry about the x1 axis.

⇒ symmetry about the x2 axis.

f ( x1 , x2 ) = − f (− x1 ,− x 2 ) ⇒ symmetry about the origin. 2.2 Constructing Phase Portraits There are a number of methods for constructing phase plane trajectories for linear or nonlinear system, such that so-called analytical method, the method of isoclines, the delta method, Lienard’s method, and Pell’s method. Analytical method There are two techniques for generating phase plane portraits analytically. Both technique lead to a functional relation between the two phase variables x1 and x2 in the form g ( x1 , x 2 ) = 0 (2.6)

where f1 , f 2 are assumed to be single valued functions. This implies that the phase trajectories will not intersect. At singular point, however, the value of the slope is 0/0, i.e., the slope is indeterminate. Many trajectories may intersect at such point, as seen from Fig. 2.2. This indeterminacy of the slope accounts for the adjective “singular”. Singular points are very important features in the phase plane. Examining the singular points can reveal a great deal of information about the properties of a system. In fact, the stability of linear systems is uniquely characterized by the nature of their singular points. Although the phase plane method is developed primarily for second-order systems, it can also be applied to the analysis of first-order systems of the form & x + f ( x) = 0 The difference now is that the phase portrait is composed of a single trajectory. Example 2.3 A first-order system_______________________

where the constant c represents the effects of initial conditions (and, possibly, of external input signals). Plotting this relation in the phase plane for different initial conditions yields a phase portrait. The first technique involves solving (2.1) for x1 and x2 as a function of time t , i.e., x1 (t ) = g1 (t ) and x2 (t ) = g 2 (t ) , and then, eliminating time t from these equations. This technique was already illustrated in example 2.1. The second technique, on the other hand, involves directly dx f (x , x ) eliminating the time variable, by noting that 2 = 2 1 2 dx1 f1 ( x1 , x 2 ) and then solving this equation for a functional relation between x1 and x2 . Let us use this technique to solve the massspring equation again. Example 2.4 Mass-spring system_______________________

& Consider the system x = −4 x + x there are three singular
points, defined by − 4 x + x 3 = 0 , namely, x = 0, − 2, 2 . The phase portrait of the system consists of a single trajectory, and is shown in Fig. 2.3.
& x

3

unstable
-2

stable

unstable
x

0

2

& By noting that && = (dx / dx) /( dx / dt ) , we can rewrite (2.2) as x & dx 2 & & x + x = 0 . Integration of this equation yields x 2 + x 2 = x0 . dx
__________________________________________________________________________________________

Fig. 2.3 Phase trajectory of a first-order system The arrows in the figure denote the direction of motion, and whether they point toward the left or the right at a particular & point is determined by the sign of x at that point. It is seen from the phase portrait of this system that the equilibrium point x = 0 is stable, while the other two are unstable.
__________________________________________________________________________________________

Most nonlinear systems cannot be easily solved by either of the above two techniques. However, for piece-wise linear systems, an important class of nonlinear systems, this can be conveniently used, as the following example shows. Example 2.5 A satellite control system___________________
Jets
θd = 0
U -U

u

1 p

Sattellite θ&

1 p

θ

2.1.3 Symmetry in phase plane portrait

& x Let us consider the second-order dynamics (2.3): && = f ( x, x) . The slope of trajectories in the phase plane is of the form
dx2 f ( x1 , x2 ) =− & dx1 x

Fig. 2.4 Satellite control system

Fig. 2.4 shows the control system for a simple satellite model. The satellite, depicted in Fig. 2.5.a, is simply a rotational unit inertia controlled by a pair of thrusters, which can provide either a positive constant torque U (positive firing) or negative torque (negative firing). The purpose of the control system is Since symmetry of the phase portraits also implies symmetry to maintain the satellite antenna at a zero angle by of the slopes (equal in absolute value but opposite in sign), we appropriately firing the thrusters. can identify the following situations: ___________________________________________________________________________________________________________ 2 Chapter 2 Phase Plane Analysis

Applied Nonlinear Control

Nguyen Tan Tien - 2002.3

_________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

& The mathematical model of the satellite is θ& = u , where u is the torque provided by the thrusters and θ is the satellite angle. Let us examine on the phase plane the behavior of the control system when the thrusters are fired according to the control law −U u (t ) =  U if if

The method of isoclines (ñöôø ng ñaú ng khuynh) The basic idea in this method is that of isoclines. Consider the & & dynamics in (2.1): x1 = f1 ( x1 , x 2 ) and x2 = f 2 ( x1 , x 2 ) . At a point ( x1 , x2 ) in the phase plane, the slope of the tangent to the trajectory can be determined by (2.5). An isocline is defined to be the locus of the points with a given tangent slope. An isocline with slope α is thus defined to be dx2 f (x , x ) = 2 1 2 =α dx1 f1 ( x1 , x 2 ) This is to say that points on the curve f 2 ( x1 , x 2 ) = α f1 ( x1 , x2 ) all have the same tangent slope α . In the method of isoclines, the phase portrait of a system is generated in two steps. In the first step, a field of directions of tangents to the trajectories is obtained. In the second step, phase plane trajectories are formed from the field of directions. Let us explain the isocline method on the mass-spring system in (2.2): && + x = 0 . The slope of the trajectories is easily seen x to be dx2 x =− 1 dx1 x2

θ >0 θ <0

(2.7)

which means that the thrusters push in the counterclockwise direction if θ is positive, and vice versa. As the first step of the phase portrait generation, let us consider the phase portrait when the thrusters provide a & positive torque U . The dynamics of the system is θ& = U , & dθ& = U dθ . Therefore, the phase portrait which implies that θ trajectories are a family of parabolas defined by θ& 2 = 2U θ + c1 , where c1 is constant. The corresponding phase portrait of the system is shown in Fig. 2.5.b. When the thrusters provide a negative torque −U , the phase trajectories are similarly found to be θ& 2 = −2U x + c1 , with the corresponding phase portrait as shown in Fig. 2.5.c.

θ

antenna

& x
x

& x
x

Therefore, the isocline equation for a slope α is x1 + α x 2 = 0

u

u =U

u = −U

Fig. 2.5 Satellite control using on-off thrusters The complete phase portrait of the closed-loop control system can be obtained simply by connecting the trajectories on the left half of the phase plane in 2.5.b with those on the right half of the phase plane in 2.5.c, as shown in Fig. 2.6. parabolic trajectories
& x

i.e., a straight line. Along the line, we can draw a lot of short line segments with slope α . By taking α to be different values, a set of isoclines can be drawn, and a field of directions of tangents to trajectories are generated, as shown in Fig. 2.7. To obtain trajectories from the field of directions, we assume that the tangent slopes are locally constant. Therefore, a trajectory starting from any point in the plane can be found by connecting a sequence of line segments.

α =1

& x

α = −1

x

x

α =∞
u = +U u = −U switching line

Fig.2.6 Complete phase portrait of the control system The vertical axis represents a switching line, because the control input and thus the phase trajectories are switched on that line. It is interesting to see that, starting from a nonzero initial angle, the satellite will oscillate in periodic motions under the action of the jets. One can concludes from this phase portrait that the system is marginally stable, similarly to the mass-spring system in Example 2.1. Convergence of the system to the zero angle can be obtained by adding rate feedback.

Fig. 2.7 Isoclines for the mass-spring system Example 2.6 The Van der Pol equation__________________ For the Van der Pol equation && + 0.2( x 2 − 1) x + x = 0 & x an isocline of slope α is defined by

& & dx 0.2( x 2 − 1) x + x =− =α & dx x __________________________________________________________________________________________ ___________________________________________________________________________________________________________ 3 Chapter 2 Phase Plane Analysis

namely the origin. if we plot a phase plane portrait & with new coordinates x and (1 / x ) . & t − t 0 ≈ (1 / x) dx x0 For linear systems described by (2. we should divide the corresponding part of the trajectory into a number of small segments (not necessarily equally spaced). & Obtaining time from ∆t ≈ ∆x / x In a short time ∆t . λ2 are complex conjugates with real parts equal to 0 We now briefly discuss each of the above four cases Stable or unstable node (Fig. The phase portraits can be obtained. 2. the length of time corresponding to the & increment ∆x is ∆t ≈ ∆x / x .10).2002. as plot in Fig.9a) and then substitution of & (2.a -b) The first case corresponds to a node. λ2 < 0 : singularity point is called stable node. α = 0 α = −1 x2 where x corresponding to time t and x0 corresponding to time t 0 . the points on the curve & & 0.8) whre the constant λ1 . almost all of the system trajectories diverge to infinity.5. we solve for the time history x(t ) = k1e λ1 t + k 2 e λ2 t x(t ) = k1e λ1 t for λ1 ≠ λ2 for λ1 = λ2 (2. By taking α of different isoclines can be obtained. We now to describe two techniques for computing time history from phase portrait. Therefore. The following cases can occur • λ1 . 2.c) The second case ( λ1 < 0 < λ2 ) corresponds to a saddle point.4 Phase Plane Analysis of Linear Systems The general form of a linear second-order system is & x1 = a x1 + b x2 & x2 = c x1 + d x 2 (2. This closed curve corresponds to a limit cycle. This implies that. This implies that. A node can be stable or unstable: λ1 .3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Therfore. ∫ x ___________________________________________________________________________________________________________ 4 Chapter 2 Phase Plane Analysis . and then plots x with respects to t and x with respects to t .11a) (2. From (2. find the time associated with each segment. λ2 are complex conjugates with non-zero real parts • λ1 . Because of the unstable pole λ2 . then the area under the resulting curve is the corresponding time interval. differentiation of (2. 2. and then add up the results. ∫ λ1 . in order to obtain the time corresponding to the motion from one point to another point along the trajectory. we can write dt = dx / x .10) isoclines Fig.3 Determining Time from Phase Portraits Time t does not explicitly appear in the phase plane having x1 and x 2 as coordinates. we simply compute the time t for each point on the & phase trajectory. It is interesting to note that there exists a closed curved in the portrait. & Obtaining time from t ≈ (1 / x) dx & & Since x = dx / dt . λ2 are the solutions of the characteristic equation s 2 + as + b = ( s − λ1 )( s − λ2 ) = 0 The roots λ1 .9a) (2.11b) + k2 t e λ2 t 2. Both of techniques involve a step-by step procedure for recovering time. the change of x is approximately & ∆x ≈ x ∆t (2. 2. Therefore. as will be discussed further in section 2.9. we x will simply consider the second-order linear system described by && + a x + b x = 0 & x (2. λ2 > 0 : singularity point is called unstable node.8. as shown in the plot. There is no oscillation in the trajectories.8). However. and the trajectories starting from both outside and inside converge to this curve.Applied Nonlinear Control Nguyen Tan Tien . there is only one singular point (b ≠ 0) .9b) α =1 trajectory -2 2 α = −5 x1 limit cycle Transform these equations into a scalar second-order & & differential equation in the form b x 2 = b c x1 + d ( x1 − a x1 ) . the trajectories in the vicinity of this singularity point can display quite different characteristics. Saddle point (Fig.2( x 2 − 1) x + x + α x = 0 all have the same slope α . depending on the values of a and b . λ2 are both real and have opposite sign • λ1 .9. To obtain the history of states corresponding to a certain initial condition. __________________________________________________________________________________________ To obtain the phase portrait of this linear system. 2. λ2 can be explicitly represented as λ1 = − a + a 2 − 4b − a − a 2 − 4b and λ2 = 2 2 & where x is the velocity corresponding to the increment ∆x . λ2 are both real and have the same sign (+ or -) • λ1 . Consequently.8 Phase portrait of the Van der Pol equation Short line segments are drawn on the isoclines to generate a field of tangent directions.9b) leads to &&1 = (a + d ) x1 + (c b − a d ) x1 .

9.10 Stable.10. 2. Therefore. Depending on the motion patterns of the trajectories in the vicinity of the limit cycle.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ jω stable node & x x σ (a) jω Local behavior of nonlinear systems If the singular point of interest is not at the origin.1) can be rewritten in the form & x1 = a x1 + b x 2 + g1 ( x1 . x 2 ) & x2 = c x1 + d x 2 + g 2 ( x1 . such as multiple equilibrium 2 2  x 2 = − x1 − x2 ( x1 + x2 − 1) 2 & points and limit cycles. diverging x1 x1 x1 limit cycle limit cycle limit cycle (c) ⊗ Note that the stability characteristics of linear systems are uniquely determined by the nature of their singularity points. and isolated.5 Phase Plane Analysis of Nonlinear Systems (a ) (b) Fig. we can shift the singular point to the origin.1) with a singular point at 0. we can distinguish three kinds of limit cycles. 2.a). The trajectory has to be both closed.7 Stable. • Stable Limit Cycles: all trajectories in the vicinity of the limit cycle converge to it as t → ∞ (Fig. the higher order terms can be neglected. unstable. ___________________________________________________________________________________________________________ 5 Chapter 2 Phase Plane Analysis . the local behavior of the nonlinear system can be approximated by the patterns shown in Fig. 2.13) 2 2 nonlinear systems can be approximated by the behavior  x2 = − x1 + x 2 ( x1 + x2 − 1) & of a linear system. λ2 ) < 0 : stable focus Re(λ1 . 2.9 Phase-portraits of linear systems Stable or unstable locus (Fig. Re(λ1 . indicating the periodic nature of the motion.f) The last case corresponds to a certain point. x jω stable focus & x σ (d ) jω unstable focus & x x σ (e) Limit cycle In the phase plane.d-e) The third case corresponds to a focus. and semi-stable limit cycles Example 2. and semi-stable limit cycle___ Consider the following nonlinear systems 2 2  x = x2 − x1 ( x1 + x 2 − 1) & (a)  1 2 2  x 2 = − x1 − x 2 ( x1 + x 2 − 1) & (2.12) In discussing the phase plane analysis of nonlinear system. we may simply consider Eq.9.9.(2. and therefore.b) • Semi-Stable Limit Cycles: some of the trajectories in the vicinity of the limit cycle converge to it as t → ∞ (Fig. (2. • Unstable Limit Cycles: all trajectories in the vicinity of the limit cycle diverge to it as t → ∞ (Fig. by defining the difference between the original state and the singular point as a new set of state variables.10.2002. 2.14) patterns in the phase plane. All trajectories are ellipses and the singularity point is the centre of these ellipses. a limit cycle is defied as an isolated closed curve. unstable. indicating the limiting nature of the cycle (with near by trajectories converging or diverging from it). is not true for nonlinear systems. however. 2. 2 2  x1 = x 2 − x1 ( x1 + x 2 − 1) 2 • Nonlinear systems can display much more complicated & (c)  (2. Using Taylor expansion. λ2 ) > 0 : unstable focus Center point (Fig. 2. 2. without loss of generality. x 2 ) where g1 .Applied Nonlinear Control Nguyen Tan Tien . unstable node & x x σ (b) jω saddle point & x x σ (c ) In the vicinity of the origin. two points should be kept in mind: • Phase plane analysis of nonlinear systems is related to 2 2  x = x 2 + x1 ( x1 + x 2 − 1) & that of liner systems. 2. g 2 contain higher order terms. This.10. the nonlinear system trajectories essentially satisfy the linearized equation & x1 = a x1 + b x2 & x2 = c x1 + d x 2 As a result. Eqs. because the local behavior of (b)  1 (2.c) x2 converging trajectories x2 diverging trajectories converging x2 jω center point & x x σ (f) Fig.

which is always strictly ∂x1 ∂x 2 positive (except at the origin). then r < 0 . __________________________________________________________________________________________ 2. & When r > 1 . Therefore. Where.12) r (t ) = 1 1 + c0 e − 2t and θ (t ) = θ 0 − t .1). This implies that the state tends to the circle from inside. then one of the following is true: (a) the trajectory goes to an equilibrium point (b) the trajectory tends to an asymptotically stable limit cycle (c) the trajectory is itself a limit cycle Theorem 2.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ By introducing a polar coordinates 2 2 r = x1 + x2 x  θ (t ) = tan −1  2  x   1 the dynamics of (2. the system does not have any limit cycles any where in the phase plane. This can also be concluded by examining the analytical solution of (2. Therefore. This implies that the states tend to the unit circle from outside.8________________________________________ Consider the nonlinear system 2 & x1 = g ( x 2 ) + 4 x1 x 2 2 & x2 = h( x1 ) + 4 x1 x2 Since ∂ f1 ∂ f 2 2 2 + = 4( x1 + x 2 ) .2002. This theorem is sometime called index theorem. where c0 = 1 r02 −1 Similarly.6 Existence of Limit Cycles Theorem 2. no limit cycle can exist in the region Ω of the phase plane in which ∂f1 / ∂x1 + ∂f 2 / ∂x 2 does not vanish and does not change sign.1). centers. then r > 0 . the unit circle is a stable limit cycle. S represents the number of enclosed saddle points.2 (Pointcare-Bendixson) If a trajectory of the second-order autonomous system remains in a finite region Ω .12) are transformed as dr = − r (r 2 − 1) dt dθ = −1 dt When the state starts on the unicycle. N represents the number of nodes. __________________________________________________________________________________________ ___________________________________________________________________________________________________________ 6 Chapter 2 Phase Plane Analysis .Applied Nonlinear Control Nguyen Tan Tien . Example 2.1 (Pointcare) If a limit cycle exists in the secondorder autonomous system (2. the N=S+1. When r < 1 . and foci enclosed by a limit cycle. the state will circle around the & origin with a period 1 / 2π . Theorem 2. the above equation & shows that r (t ) = 0 .3 (Bendixson) For a nonlinear system (2. we can find that the system (b) has an unstable limit cycle and system (c) has a semi-stable limit cycle.

A linear time-invariant system & x = Ax * (3. for notational and analytical simplicity. and study the associated variation of the motion error e(t ) = x(t ) − x* (t ) as illustrated in Fig.1. A special class of nonlinear systems is linear system. b the friction coefficient at the hinge. 3.2002.. 3. and g the gravity constant. Nominal motion & Let x* (t ) be the solution of x = f (x) .2) In linear system analysis and design. this means that the constant vector x * satisfies 0 = f (x ) Equilibrium points can be found using (3.1) θ R : nonlinear vector function : state vectors : order of the system Fig. t ) where f ∈R n has a single equilibrium point (the origin 0) if A is nonsingular. A nonlinear system can have several (or infinitely many) isolated equilibrium points.6a) (3. which leads to the points (0 [2π ]. 0) and (π [2π ]. The & dynamics of linear systems are of the from x = A(t ) x with A ∈ R n×n . if the system’s state equation can be written & x = f (x) Otherwise. 3. Such a point is called an equilibrium point. As we shall see later. the system is called non-autonomous. it has an infinity of equilibrium points. Let us now perturb the initial condition to be x(0) = x 0 + δ x 0 . Definition 3. many stability problems are naturally formulated with respect to equilibrium points.6b) Therefore the equilibrium points are given by x2 = 0. Definition 3. 3.1 Pendulum Consider the pendulum of Fig.. Equilibrium points It is possible for a system trajectory to only a single point.1) can represent both closed-loop dynamics of a feedback control system and the dynamic systems where no control signals are involved. Fundamentals of Lyapunov Theory The objective of this chapter is to present Lyapunov stability theorem and illustrate its use in the analysis and the design of nonlinear systems. For nonlinear systems. i. i. __________________________________________________________________________________________ (3.e.1 The nonlinear system (3. the nominal motion trajectory. these adjectives are replaced by autonomous and non-autonomous. the corresponding state-space equation is & x1 = x2 g & x2 = − x − sin x1 2 2 R MR b (3.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 3. which contained in the null-space of the matrix A. i. these points correspond to the pendulum resting exactly at the vertical up and down points. Leting x1 = θ . the subspace defined by Ax = 0.2 A state x is an equilibrium state (or * where R is the pendulum’s length.1 The pendulum___________________________ (3. whose dynamics is given by the following nonlinear autonomous equation & MR 2θ& + bθ& + MgR sin θ = 0 (3.. Example 3.1) is said to be autonomous if f does not depend explicitly on time. 3.2 Nominal and perturbed motions ___________________________________________________________________________________________________________ 7 Chapter 3 Fundamentals of Lyapunov Theory (3. Mathematically.2. x2 = θ& .1 Nonlinear Systems and Equilibrium Points Nonlinear systems A nonlinear dynamic system can usually be presented by the set of nonlinear differential equations in the form & x = f ( x.4) .e. If A is singular. x2 x(t ) e(t ) equilibrium points) of the system if once x(t ) is equal to x * . Autonomous and non-autonomous systems Linear systems are classified as either time-varying or timeinvariant. sin( x1 ) = 0. we often transform the linear system equations in such a way that the equilibrium point is the origin of the state-space.5) x ∈ Rn n The form (3.e.3). corresponding to initial condition x* (0) = x 0 .Applied Nonlinear Control Nguyen Tan Tien . it remains equal to x * for all future time. M its mass. 0) . Physically.3) x* (t ) x1 xn Fig.

9) means that the state vector of an exponentially stable form system converges to the origin faster than an exponential ___________________________________________________________________________________________________________ 8 Chapter 3 Fundamentals of Lyapunov Theory . there exist r > 0 . The ball B r is called a domain of attraction of the equilibrium point. Lyapunov stability is not enough. the equilibrium point is unstable.3 can be written in the (3. Example 3. More formally. instead of studying the deviation of x(t ) from x * (t ) for the original system. Therefore. has an equilibrium point at the origin of the state space. 3.8) with initial condition e(0) = δ x(0) . Using the above symbols. or Lyapunov stability) means that the system trajectory can be kept arbitrarily close to the origin by starting sufficiently close to it. t ) (3. and if in addition there exist some r > 0 such that x(0) ≤ r implies that x(t ) → 0 as t → ∞ .2): x = f (x) . but also required that the attitude gradually go back to its original value. i.asymptotically stable 3 1 curve 2 . Assume that we slightly perturb the initial position to be x(0) = x0 + δ x0 . the equivalent differential equation governing the motion error e is m && + k1e + k 2 [e 3 + 3e 2 x * (t ) + 3e x *2 (t )] = 0 e Clearly.8) with respect to the equilibrium point 0.9) x* (t ) which starts from initial point x0 . the perturbation dynamics non-autonomous system. we not only want the satellite to maintain its attitude in a range determined by the magnitude of the disturbance.. Otherwise. given that we do not want the state trajectory x(t ) to get out of a ball of arbitrarily specified radius B R .3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ & Since both x * (t ) and x(t ) are solutions of (3. For example. Lyapunov stability. we may simply study the stability of the perturbation dynamics (3. This type of engineering requirement is captured by the concept of asymptotic stability. x(t ) ≤ α x(0) e −λt in some ball B r around the origin. t ) − f (x * . The resulting system trajectory is denoted as x(t ) . and in addition. Fig. Let us study the stability of the motion Essentially. stability (also called stability in the sense of Lyapunov. Definition 3. equivalently ∀R > 0. curve 1 . the definition states that the origin is stable. we have & x * (0) = x 0 x* = f (x* ) ∀R > 0.2 Concepts of Stability Notation B R : spherical region (or ball) defined by x ≤ R S R : spherical itself defined by x = R ∀ ∃ ∈ ⇒ ⇔ : for any : there exist : in the set : implies that : equivalent Stability and instability Definition 3. states start close to 0 actually converge to 0 as time goes to infinity. for any R > 0 .3 shows that the system trajectories starting form within the ball B r converge to the origin. if.unstable SR Fig.4 An equilibrium points 0 is asymptotically stable if it is stable. 3.marginally stable 2 0 x(0) S r curve 3 .2002. due to the presence of the nominal trajectory x * (t ) on the right hand side. x(t ) < R or. Definition 3. Definition 3. There is a need to estimate how fast the system trajectory approaches 0. ∃r > 0. However.5 An equilibrium points 0 is exponential stable if there exist two strictly positive number α and λ such that ∀t > 0. x(0) ∈ B r ⇒ ∀t ≥ 0.e. x(t ) ∈ B r & x = f (x) x ( 0) = x 0 + δ x 0 then e(t ) satisfies the following non-autonomous differential equation & e = f (x * + e. The geometrical implication of stability is indicated in Fig.3 Concepts of stability Asymptotic stability and exponential stability In many engineering applications. such that if x(0) ≤ r then x(t ) ≤ R for all t ≥ 0 . t ) = g (e. when a satellite’s attitude is disturbed from its nominal position. __________________________________________________________________________________________ 3.Applied Nonlinear Control Nguyen Tan Tien .3 The equilibrium state x = 0 is said to be stable if. 33. x(0) < r ⇒ ∀t ≥ 0. Proceeding as before. The concept of exponential stability can be used for this purpose. this is a non-autonomous system. (3. In many engineering applications. ∃r > 0. t ) = 0 . with e as state and g in place of f.2________________________________________ Consider the autonomous mass-spring system & m && + k1 x + k 2 x 3 = 0 x which contains a nonlinear term reflecting the hardening effect of the spring. Since g (0. Asymptotic stability means that the equilibrium is stable. it is still not sufficient to know that a system will converge to the equilibrium point after infinite time. the new dynamic system.

finding a system’s linearization is often most easily done simply neglecting any term of order higher than 1 in the dynamics. if all eigenvalues of A are in the left-half complex plane but at least one of them is on the jω axis). = π − θ + h. stands for higher-order terms in x. as we now illustrate. x(0) = 1 (3. asymptotically stable.2). and therefore x(t ) ≤ x(0) e −t .3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ function. (x)    x =0 (3. Then the system dynamics can be written as  ∂f  & x=  ∂ x  x + f h. as we shall see. and therefore so is the nonlinear system at this equilibrium point. Then. if all eigenvalues of A are strictly in the left-half complex plane).2002. a function slower than any exponential function e − λt . 3. then the equilibrium point is asymptotically stable (for the actual nonlinear system). or unstable for the nonlinear system).4________________________________________ Consider the nonlinear system 2 & x1 = x2 + x1 cos x 2 & x2 = x 2 + ( x1 + 1) x1 + x1 sin x 2 Hence its linear approximation is unstable.5________________________________________ Consider the equilibrium point (θ = π .5________________________________________ & Consider the first-order system x = a x + b x 5 . __________________________________________________________________________________________ 1 1 ___________________________________________________________________________________________________________ 9 Chapter 3 Fundamentals of Lyapunov Theory .e. and assumed that f(x) is continuously differentiable. then one cannot conclude anything from the linear approximation (the equilibrium point may be stable. if at least one eigenvalue of A is strictly in the right-half complex plane). the nonlinear system is x = b x 5 .o. • If the linearizad system is un stable (i. the equilibrium point is said to be asymptotically (or exponentially) stable in the large.t.o.. whose solution is x = 1 /(1 + t ) .t .o. & Consider the system && + 4 x 5 + ( x 2 + 1) u = 0 . we can write sin θ = sin π + cos π (θ − π ) + h.Applied Nonlinear Control Nguyen Tan Tien . the system    x =0 & x = Ax (3.t . But asymptotic stability does not implies guarantee exponential stability. It is also called globally asymptotically (or exponentially) stable. The linearization of & this system around the origin is x = a x .o. __________________________________________________________________________________________ Example 3. Local and global stability Definition 3. Theorem 3. The linearization method fails while. The origin 0 is one of the two equilibrium of this system. Let us use the constant matrix A denote the Jacobian matrix of f with respect  ∂f  to x at x = 0: A =   ∂ x  . Example 3. Note that exponential stability implies asymptotic stability.θ& = 0) is ~ && θ + ~ g ~ & θ − θ =0 R MR b 2 where f h.2).3 Linearization and Local Stability Lyapunov’s linearization method is concerned with the local stability of a nonlinear system. The positive number λ is called the rate of exponential convergence. Consider the autonomous system in (3. as can be seen from the system & x = − x 2 . Indeed. its solution is x(t ) = x(0) e t − ∫0 −[1+sin 2 x (τ )] dτ .11) • If the linearized system is marginally stable (i.10) __________________________________________________________________________________________ The following result makes precise the relationship between the stability of the linear system (3.6 If asymptotic (or exponential) stability holds for any initial states.e. then the equilibrium point is unstablle (for the nonlinear system). Since the neighborhood of θ = π .1 (Lyapunov’s linearization method) • If the linearized system is strictly stable (i. A similar procedure can be applied for a controlled system. x & & For example. Assume that the control law for the original nonlinear x system has been selected to be u = sin x + x 3 + x cos 2 x .. the system’s linearization about the equilibrium point (θ = π . The system can x be linearly approximated about x = 0 as && + 0 + (0 + 1) u = 0 or x && = u .θ& = 0) of the pendulum in the example 3.t. It is a formalization of the intuition that a nonlinear system should behave similarly to its linearized approximation for small range motions..2) and that of the original nonlinear system (3.1. Example 3. The application of Lyapunov’s linearization method indicate the following stability properties of the nonlinear system • a < 0 : asymptotically stable • a > 0 : unstable • a = 0 : cannot tell from the linearization & In the third case. the system x = −(1 + sin 2 x) x is exponentially convergent to x = 0 with the rate λ = 1 . the direct method to be described can easily solve this problem.1 & x2 = x 2 + 0 + x1 + x1 x2 ≈ x2 + x1 1 0 & The linearized system can thus be written x =  x . In practice.12) is called the linearization (or linear approximation) of the original system at the equilibrium point 0. ~ thus letting θ = θ − π .e. then the linearized closed-loop dynamics is && + x + x = 0 . Its linearized approximation about x = 0 is & x1 = 0 + x1.

2) is & negative semi-definite.e. The direct method of Lyapunov is based on generalization of the concepts in the above mass-spring-damper system to more complex systems. (3.7 A scalar continuous function V (x) is said to be locally positive definite if V (0) = 0 and.6. must eventually settle down to an equilibrium point. starting from some initial value.1. 3.4 Lyapunov’s Direct Method The basic philosophy of Lyapunov’s direct method is the mathematical extension of a fundamental physical observation: if the total energy of a mechanical (or electrical) system is continuous dissipated. and if its time derivative along any state trajectory of system (3. x 2 ) = Vα typically present a set of ovals surrounding the origin. V (x) typically corresponds to a surface looking like an upward cup as shown in Fig.15) implies that the energy of the system.These ovals often called contour curves may be thought as the section of the cup by horizontal planes. x = 0 . is continuously dissipated by the damper & until the mass is settled down. the function V (x) = 1 2 MR 2 x2 + MR(1 − cos x1 ) 2 which is the mechanical energy of the pendulum in Example 3. we may conclude the stability of a system by examining the variation of a single scalar function. 3.14) Comparing the definitions of stability and mechanical energy. V (x) is said to be a Lyapunov function for the system (3. is locally positive definite. that the stability properties of the system can be characterized by the variation of the mechanical energy of the system. indirectly reflects the magnitude of the state vector. i. then V (x) is said to be globally positive definite.14) and using (3. The rate of energy variation during the system’s motion is obtained by differentiating the first equality in (3. Let consider the nonlinear mass-damper-spring system in Fig. the mechanical energy.2002. in a ball B R0 x≠0 ⇒ V ( x) > 0 If V (0) = 0 and the above property holds over the whole state space. 3. The lowest point of the cup is located at the origin.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 3. x 2 ) The 2-dimesional geometrical representation can be made as follows.6 A nonlinear mass-damper-spring system Total mechanical energy = kinetic energy + potential energy 1 1 k 0 x 2 + k1 x 4 2 4 0 (3. i. Positive definite functions and Lyapunov functions Definition 3.13) & &x & & & & & V (x) = m x && + (k 0 x + k1 x 3 ) x = x (−b x x ) = −b x 3 (3.15) Fig. with each oval corresponding to a positive value of Vα ..e. Taking x1 and x2 as Cartesian coordinates.7 Typical shape of a positive definite function V ( x1 . & & m && + b x x + k 0 x + k1 x 3 = 0 x with b (3.. Consider a positive definite function V (x) of two state variables x1 and x2 .8 Interpreting positive definite functions using contour curves Definition 3.Applied Nonlinear Control Nguyen Tan Tien . x = 0) • assymptotic stability implies the convergence of mechanical energy to zero • instability is related to the growth of mechanical energy The relations indicate that the value of a scalar quantity. x 2 ) plane as shown in Fig. 3. Thus. whose dynamic equation is 3. 3.2). we can see some relations between the mechanical energy and the concepts described earlier: V ( x) = 1 2 & mx + 2 V = V1 x2 x1 0 ∫ (k x + k x )dx = 2 mx& 0 1 3 x 1 2 + V3 > V2 > V1 Fig. 3. the function V (x) is positive definite and has continuous partial derivatives. whether linear or nonlinear. V k 0 x + k1 x 3 : nonlinear spring term nonlinear spring and damper m V = V3 V = V2 Fig. V (x) ≤ 0 then. then the system. For instance.4.13) & x & x : nonlinear dissipation or damping Let us describe the geometrical meaning of locally positive definite functions. in a ball B R0 .1. projected on the ( x1 .8. ___________________________________________________________________________________________________________ 10 Chapter 3 Fundamentals of Lyapunov Theory . and furthermore.7. In 3-dimensional space. V = V2 x2 V = V1 x1 0 V = V3 V3 > V2 > V1 • zero energy corresponds to the equilibrium point & (x = 0. the level curves V ( x1 .8 If.

3. this function represents the total energy of the pendulum. but is otherwise arbitrary. & because V (x) is only negative semi-definite.3 (Global Stability) Assume that there exists a scalar function V of the state x. In applying the above theorem for analysis of a nonlinear system.10. this function is locally positive definite. composed of the sum of the potential energy and the kinetic energy. In Fig. in the region defined by ( x1 + x2 ) < 2 . Actually. actually. with this Lyapunov function. i. 3. 2 2 V ( x1 . ___________________________________________________________________________________________________________ 11 Chapter 3 Fundamentals of Lyapunov Theory .4. __________________________________________________________________________________________ Lyapunov theorem for global stability Theorem 3. & V (x) is precisely the power dissipated in the pendulum. then the stability is asymptotic. in a ball B R0 .10 Illustrating Definition 3.. we cannot draw conclusion on the asymptotic stability of the system.e. this condition indicates that − c(x ) ’pushes’ the system back towards its rest position x = 0 .2002. we must go through two steps: choosing a positive Lyapunov function.2 Equilibrium point theorems Lyapunov’s theorem for local stability Theorem 3. such as x c( x) > 0 ∀x ≠ 0 . 3. However. Example 3. the above theorem indicates that the origin is asymptotically stable.e. x2 ) is seen always point down an inverted cup.8 for n=2 using contour curves 3. by involving the above theorem. Example 3.7 Local stability___________________________ A simple pendulum with viscous damping is described as & θ& + θ& + sin θ = 0 Consider the following scalar function 1 V (x) = (1 − cosθ ) + θ& 2 2 & its derivative V along any system trajectory is 2 2 2 2 & V = 2( x1 + x 2 )( x1 + x 2 − 2) Thus. Intuitively. Therefore.. with continuous first order derivatives such that • V (x) is positive definite & • V (x) is negative definite • V (x) → ∞ as x → ∞ then the equilibrium at the origin is globally asymptotically stable. we can see the reason why V (x) ≤ 0 . using & physical meaning.8 Asymptotic stability_______________________ Fig. and then determining its derivative along the path of the nonlinear systems.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ A Lyapunov function can be given simple geometrical interpretations. is locally negative definite in the 2-dimensional ball B 2 .9 Illustrating Definition 3.9. __________________________________________________________________________________________ V 0 x2 x(t ) x1 Example 3.Applied Nonlinear Control Nguyen Tan Tien . we can conclude that the origin is a stable equilibrium point. If. x 2 ) = x1 + x2 Fig. Its time derivative yields & & V (x) = θ& sin θ + θ&θ& = −θ& 2 ≤ 0 Therefore. 2 2 i.9 A class of first-order systems_______________ Consider the nonlinear system & x + c( x) = 0 where c is any continuous function of the same sign as its scalar argument x . In fact. there exists a scalar function V (x) with continuous first partial derivatives such that • V (x) is positive definite (locally in B R0 ) & • V (x) is negative semi-definite (locally in B R0 ) then the equilibrium point 0 is stable. 3. namely that the damping term absorbs energy. V Obviously. the point denoting the value of V ( x1 . the state point is seen to move across contour curves corresponding to lower and lower value of V . In Fig.2 (Local stability) If.8 for n=2 V = V2 V = V1 x1 0 V = V3 V3 > V2 > V1 Let us study the stability of the nonlinear system defined by 2 2 2 & x1 = x1 ( x1 + x 2 − 2) − 4 x1 x 2 2 2 2 & x2 = 4 x1 x2 + x 2 ( x1 + x2 − 2) x2 around its equilibrium point at the origin. the & derivative V (x) is locally negative definite in B R0 . As a mater of fact.

Then. Assume that • for some l > 0 . Its derivative & & & is V = 2 x x = −2 x c ( x) .15).3 Invariant set theorem Definition 3. nor is the set M.M is the union of all invariant sets (e. x1 Fig. the theorems in Lyapunov analysis are all sufficiency theorems. with f continuous.4. Using the invariant set theorem. A precise statement of this result is as follows. the region Ω l defined by V (x) < l is bounded & • V (x) ≤ 0 for all x in Ω l & Let R be the set of all points within Ω where V (x) = 0 . Example 3 . Similarly.e.Applied Nonlinear Control Nguyen Tan Tien .3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Since c is continuous. the & system x = − x 3 is globally asymptotically convergent to x = 0 . it also implies that c(0) = 0 (Fig.Along the same line. __________________________________________________________________________________________ M be the largest invariant set in R. 3. equilibrium points or limit cycles) within R .g. 2 2 & Its derivative along any system trajectory is V = −2( x1 + x 2 ) 2 which is negative definite. . we cannot draw any conclusions on the stability or instability of the system – the only conclusion we should draw is that a different Lyapunov function candidate should be tried.e. __________________________________________________________________________________________ Ωl R M x0 x2 ⊗ Note that: .4 (Local Invariant Set Theorem) Consider an autonomous system of the form (3. The asymptotic stability result in the local Lyapunov theorem can be viewed a special case of the above invariant set theorem.13 The function c(x ) & For instance. so that x = 0 is a globally asymptotically stable equilibrium point.In particular. Let V be the positive definite function V = x1 + x2 . 3.. Thus V < 0 as long as x ≠ 0 .11______________________________________ Asymptotic stability of the mass-damper-spring system For the system (3. . the condition on V are not met. where a trajectory starting from within the bounded region Ω l is seen to converge to the largest invariant set M.. if the set R is itself invariant (i. the actual nonlinear system enjoys a strong stability property (global asymptotic stability). even about local stability.13). we can only draw conclusion of marginal stability using the energy function (3. specific choices of Lyapunov functions may yield more precise results than others. every solution x(t ) originating in Ω l tends to M as t → ∞ .Many Lyapunov function may exist for the same system. Note that the globalness of this stability result also implies that the origin is the only equilibrium point of the system.14) in the local & equilibrium point theorem. ⊗ Note that: . ___________________________________________________________________________________________________________ 12 Chapter 3 Fundamentals of Lyapunov Theory . Consider as the Lyapunov function candidate the square of distance to the origin V = x 2 .2002. sin 2 x ≤ sin x ≤ x . the origin is a globally asymptotically stable equilibrium point. then ≡ 0 for all future time). 3.13). and let V (x) be a scalar function with continuous first partial derivatives. the system x = sin 2 x − x is globally convergent to x = 0 . The function V is radially unbounded. we only have to show that the set M contains only one point. TO do this.14. ) V has to converge to zero) because V is lower bounded. because V is only negative semidefinite according to (3.. if once & V = 0 . then M=R The geometrical meaning of the theorem is illustrated in Fig. and l 0 x Fig. V =l V Example 3 . c (x ) Local invariant set theorem The invariant set theorem reflect the intuition that the decrease & of a Lyapunov function V has to graduate vanish (i. If for a particular choice of & Lyapunov function candidate V . we can show that the system is actually asymptotically stable.2). 3.14 Convergence to the largest invariant set M Let us illustrate applications of the invariant set theorem using some examples. & Notice that while this system’s linear approximation ( x ≈ 0) is inconclusive.9 A set G is an invariant set for a dynamic system if every system trajectory which starts from a point in G remains in G for all future time.For a given system. Theorem 3. since for x ≠ 0 . 3. since it tends to infinity as x → ∞ . Therefore. Note that the set R is not necessarily connected. where the set M consists only of the origin. however.10______________________________________ Consider the nonlinear system 2 2 & x1 = x2 − x1 ( x1 + x2 ) 2 2 & x2 = − x1 − x 2 ( x1 + x2 ) The origin of the state-space is an equilibrium point for this 2 2 system.

the 2 2 region Ω l . Actually. with f continuous. all system trajectories starting in Ω l converge either to the limit cycle or the origin (Fig. 3. actually converges to the limit cycle.e. For any arbitrary positive number l . __________________________________________________________________________________________ Example 3 . Let us show that the largest invariant set M in this set R contains only the origin. Its derivative 10 6 4 2 & V = −8( x1 + 3x 2 )( x1 + 2 x2 − 10) 2 4 2 & Thus V is strictly negative. __________________________________________________________________________________________ Fig.Applied Nonlinear Control Nguyen Tan Tien . except if x1 + 2 x 2 = 10 or 10 6 & x1 + 3 x 2 = 0 .The set Ω itself is not necessarily a domain of attraction. a contradiction to the definition. The set R is simply the origin 0. any trajectory starting within the circle converges to the origin. the set M simply consists of their union.8. a domain of attraction is explicitly determined by the invariant set theorem. defined by V ( x1 .2). ⊗ Note that: . combined with a third condition on the trajectories within R. is bounded. therefore. The first equation is simply that defining the limit cycle. or the whole horizontal axis in the phase & plane ( x.The largest connected region of the form Ω l within Ω is a domain of attraction of the equilibrium point. and let V (x) be a scalar function with continuous partial derivatives.13 Attractive limit cycle_____________________ Consider again the system 7 4 2 & x1 = x2 − x1 ( x1 + 2 x 2 − 10) Example 3. Thus. the above theorem does not guarantee that Ω is invariant: some trajectories starting in Ω but outside of the largest Ω l may actually end up outside Ω .The above corollary replaces the negative definiteness & condition on V in Lyapunov’s local asymptotic stability & theorem by a negative semi-definiteness condition on V . in which cases V = 0 . which surrounds the limit circle. the equilibrium point at the origin can actually be shown to be unstable. we see that the invariant set actually represents a limit circle. 3. Is this limit circle actually attractive ? Let us define a Luapunov 4 2 function candidate V = ( x1 + 2 x2 − 10) 2 which represents a measure of the “distance” to the limit circle. by enlarging the involved region to be the whole space and requiring the radial unboundedness of the scalar function V .11 actually represents a very common application of the invariant set theorem: conclude asymptotic stability of & an equilibrium point for systems with negative semi-definite V . which is an invariant set (since it is an equilibrium point). is bounded. All the conditions of the local invariant set theorem are satisfied and. Any state trajectory starting from the region within the limit cycle. The motion on this invariant set is described (equivalently) by either of the equations & x1 = x 2 3 & x 2 = − x1 Therefore. because the function V is not unique.15 Convergence to a limit circle Moreover. Since both the limit circle and the origin are invariant sets. For l = 1 . . along which the state vector moves clockwise. Furthermore. the acceleration at that point is && = −(k 0 / m) x − (k1 / m) x 3 ≠ 0 . since d 4 2 10 6 4 2 ( x1 + 2 x 2 − 10) = −(4 x1 + 12 x 2 )( x1 + 2 x2 − 10) dt which is zero on the set. excluding the origin. while the second equation is verified only at the origin. the largest connected region of the form (defined by V (x) < l ) within Ω is a domain of attraction of the equilibrium point.12 Domain of attraction____________________ Consider again the system in Example 3. x 2 ) = x1 + x2 < 1 . & x2 = 3 − x1 5 4 − 3 x 2 ( x1 2 + 2 x2 − 10) 4 2 Note that the set defined by x1 + 2 x2 = 10 is invariant. the largest invariant set M in R then contains only the equilibrium point 0.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ & The set R defined by x = 0 . then. The following corollary of the invariant set theorem is more specifically tailored to such applications. __________________________________________________________________________________________ x2 limit cycle 0 x1 Example 3 . the collection of states with zero velocity. Assume that in a certain neighborhood Ω of the origin • is locally positive definite & • V (x) is negative semi-definite & • the set R defined by V (x) = 0 contains no trajectories of (3. Indeed. the equilibrium point 0 is asymptotically stable. but not necessarily the whole domain of attraction.. Global invariant set theorem The above invariant set theorem and its corollary can be simply extended to a global result. This implies that the x trajectory will immediately move out of the set R and thus also out of the set M. . Assume that M contains a point with a non-zero position x1 .2) other than the trivial trajectory x ≡ 0 Then.15) ___________________________________________________________________________________________________________ 13 Chapter 3 Fundamentals of Lyapunov Theory .2002. the region Ω l . Thus. i. Corollary: Consider the autonomous system (3. x) .

Now x = 0 implies that && = −c (x) . 0 ∫ x then V is a radially unbounded function and the equilibrium point at the origin is globally asymptotically stable. since they correspond to local minima of V (note again that linearization is inconclusive about their stability).13 is actually global: all system trajectories converge to the limit cycle (unless they start exactly at the origin. Then all solutions globally asymptotically converge to M as t → ∞ For instance. The first two of these equilibrium points are stable. the virtual Chose the Lyapunov function V = ∫  y − sin 0 x Fig. Let us consider each of cases: πx & x=0 ⇒ && = sin x − x ≠ 0 except if x = 0 or x = ±1 2 x = ±1 ⇒ && = 0 x Thus the invariant set theorem indicates that the system & converges globally to or ( x = −1. the largest invariant set M in R & contains only one point. 3. 3.16) vC = c (x ) v L = && x & v R = b(x ) 0 & x 0 x & Fig. 3. Use of the local invariant set theorem indicates that the origin is a locally asymptotically stable point. Assume that & • V (x) ≤ 0 over the whole state space • V (x) → ∞ as x → ∞ as long as x ≠ 0 . x = 0 . A positive definite function for this system is x 1 2 & V = x + c( y ) dy . the above theorem shows that the limit cycle convergence in Example 3.17). as can be shown from linearization ( && = (π / 2 − 1) x) . according to the global invariant set theorem. Furthermore. and let V (x) be a scalar function with continuous first partial derivatives. if the integral c(r )dr is unbounded as x → ∞ .5 (Global Invariant Set Theorem) Consider an autonomous system of the form (3. the sign conditions b and c are simply that b(0) = 0 and c = 0 (Fig. and M be the largest invariant set in R.2002. x b( x) = 0 only if & = 0 . we obtain & & power “dissipated” by the system. or simply by noticing that because that point x is a local maximum of V along the x axis. and a local maximum in x (a saddle point in the state-space) at & & & x = 0. Because of the importance of this theorem. in other words. Thus the system cannot get “stuck” at an equilibrium value other than x = 0 . the above sign conditions are simply the necessary and sufficient conditions for the system’s stability (since they are equivalent to the conditions α1 > 0. x = 0) is unstable. 3. __________________________________________________________________________________________ ∫ & &x & & & & & & & V = x && + c( x) x = − x b( x) − x c( x) + c( x) x = − x b( x) ≤ 0 which can be thought of as representing the power dissipated & & in the system. 3. Differentiating V . by hypothesis. & b(x) c (x ) & Let R be the set of all points where V (x) = 0 ..2). namely [ x = 0.α 0 > 0 ).5 System Analysis Based on Lyapunov’s Direct Method How to find a Lyapunov function for a specific problem ? There is no general way of finding Lyapunov function for ___________________________________________________________________________________________________________ 14 Chapter 3 Fundamentals of Lyapunov Theory . any small deviation in the x direction will drive the trajectory away from it. __________________________________________________________________________________________ Example 3 . The dynamics of a mass-damper-spring system with nonlinear damper and spring can be described by the equation of this form.14 A class of second-order nonlinear systems___ Consider a second-order system of the form && + b( x) + c( x) = 0 & x where b and c are continuous functions verifying the sign & & & & conditions x b( x) > 0 for x ≠ 0 and x c( x) > 0 for x ≠ 0 . x = 0] . with the above sign conditions simply indicating that the otherwise arbitrary function b and c actually present “damping” and “spring” effects. which is an unstable equilibrium point). x = 0) . i. Together with the continuity assumptions. at x = ±1. the equilibrium point & ( x = 0. x = 0 .17 The functions b(x ) and c(x ) Furthermore. with f continuous. Example 3 . 2  & This function has two minima. c( x) = α x ) .e. which can be thought of as the sum of 2 0 the kinetic and potential energy of the system. A nonlinear R-L-C (resistorinductor-capacitor) electrical circuit can also be represented by the above dynamic equation (Fig.16 A nonlinear R-L-C circuit Note that if the function b and c are actually linear & & (b( x) = α1 x. Its derivative V = − x 2 − 1 x 4 .Applied Nonlinear Control Nguyen Tan Tien . which is non-zero & x x ⊗ Note that: Several Lyapunov function may exist for a given system and therefore several associated invariant sets may be derived.15 Multimodal Lyapunov Function___________ Consider the system && + x 2 − 1 x 3 + x = sin & x πx 2 1 2 & x + 2 π y  dy . Now V = 0 ⇒ x = 0 or x = ±1 . By contrast. let us present an additional (and very useful) example.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Theorem 3. with R & being the set defined by x = 0 .

Faced with specific systems. A time-varying matrix M(t) is uniformly positive definite if ∃α > 0.A necessary condition for a square matrix M to be p.6 A necessary and sufficient condition for a LTI & system x = A x to be strictly stable is that.s.. Its derivative yields & & & V = x T P x + x T P x = -xT Q x where A T P + P A = -Q (3. ∀t ≥ 0.(3.d. A square matrix M is skewsymmetric if M = −M T (i. __________________________________________________________________________________________ M = U T ΛU (3. Lyapunov functions for linear time-invariant systems & Given a linear system of the form x = A x . 3. skew-symmetric. j M ij = − M ji ).5.16) In the designing some tracking control systems for robot. even for stable systems. Theorem 3. Therefore. ∀i.d. . Example 3 .Applied Nonlinear Control Nguyen Tan Tien . intuition. Consider the second order linear system with A =  0   − 8 − 12 If we take P = I . a square n × n matrix M is said to be positive semi-definite (p. . for any symmetric p.Any square n × n matrix can be represented as the sum of a symmetric and a skew-symmetric matrix.18 ______________________________________ Consider again the second order linear system in Example p  p 3. Since x M x is a scalar. Let M be a n × n skew-symmetric matrix and x is an arbitrary n × 1 vector. matrix M can always be decomposed as T T T A more useful way of studying a given linear system using quadratic functions is.d. For instance.37) where U T U = I . The − 4 − 24   matrix Q is not p. 3. if ∀i.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ nonlinear system. we have to use experience.11 A square matrix M is positive definite (p.. M (t ) ≥ α I . . matrix Q.1 Lyapunov analysis of linear time-invariant systems Symmetric.2 Krasovskii’s method Krasovskii’s method suggests a simplest form of Lyapunov function candidate for autonomous nonlinear systems of the ___________________________________________________________________________________________________________ 15 Chapter 3 Fundamentals of Lyapunov Theory . • choose a positive definite matrix Q • solve for P from the Lyapunov equation • check whether P id p. p 21 = p12 .19) be symmetric positive definite. .e.18.19) is so-called Lyapunov equation. is that all its eigenvalues be strictly positive. In this section. matrix is invertible.d.d.d.) if ∀x. negative definite.d. this fact is very useful because it can simplify the control law. ⊗ Note that: . let us consider a & quadratic Lyapunov function candidate V = xT P x .19) (3. Definition 3. Note that Q may be not p.16) is a necessary and sufficient condition for a matrix M to be skew-symmetric. x T M x ≥ 0 .MT M= + 1 24 42 3 1 24 42 3 symmetric skew− symmetric . __________________________________________________________________________________________ ⊗ Note that: ..17 ______________________________________ 4 . p12 = p 22 = 1 . then . and therefore the linear system is 1 1   globally asymptotically stable. no conclusion can be draw from the Lyapunov function on whether the system is stable or not.10 A square matrix M is symmetric if M=MT (in other words. This can be shown in the following decomposition M + MT M . The corresponding matrix P = 5 1 is p.5.d.A p.d. The definition of skew-symmetric matrix implies that xT M x = − xT M T x .2002. Λ is a diagonal matrix containing the eigenvalues of M .d. Then the Lyapunov equation is  p11 p12   0 4  0 − 8   p11 p12  − 1 0   p 21 p 22  − 8 − 12 + 4 − 12  p 21 p 22  =  0 − 1         whose solution is p11 = 5 . Let us take Q = I and denote P by P =  11 12  . and positive definite matrices Definition 3. i.Q = P A + A T P =  0 −4  . then (1 / 2)x T P x is a Lyapunov function for the linear system.There are some following facts • λmin (M ) x 2 ≤ xT Mx ≤ λmax (M ) x 2 • x T Mx = xT U T ΛUx = z T Λz where Ux = z • λmin (M ) I ≤ Λ ≤ λmax (M ) I • zT z = x 2 The concepts of positive semi-definite.e. Example 3 . and physical insights to search for an appropriate Lyapunov function. x T M x = 0 (3. we discuss a number of techniques which can facilitate the otherwise blind of Lyapunov functions.. x M x = − x M x which yields ∀x.The quadratic function associated with a skew-symmetric matrix is always zero.d. p 21 p 22   where due to the symmetry of P. to derive a p.18) (3.A necessary and sufficient condition for a symmetric matrix M to be p. j M ij = M ji ). instead. matrix P from a given p.. the unique matrix P solution of the Lyapunov equation (3. And the global asymptotical stability is guaranteed.d. If P is p. and negative semi-definite can be defined similarly.A .) if x ≠ 0 ⇒ xT M x > 0 .d. matrix Q. where P is a given symmetric positive definite matrix.d. is that its diagonal elements be strictly positive.

0) dx2 + K + ∫ xn 0 ∇Vn ( x1 . then the system is globally asymptotically stable. instead of assuming a specific form for a Lyapunov function V itself. V ( x) = The applicability of the above theorem is limited in practice. A simple way is to assume that the gradient function is of the form ∇Vi = ∑a x j =1 n ij j (3. Then a sufficient condition for the origin to be asymptotically stable is that there exist two symmetric positive definite matrices P and Q.e. namely.K. in the case n = 2 . a Lyapunov function candidate is V (x) = (−6 x1 + 2 x 2 ) 2 3 + (2 x1 − 6 x 2 − 2 x 2 ) 2 The principle of the variable gradient method is to assume a specific form for the gradient ∇V . for systems of higher order. According to the theorem. To start with.7 (Krasovkii) Consider the autonomous system defined by (3. and let A(x) denote the Jacobian matrix of the system. j = 1. V = f T f . 3. the origin is asymptotically stable.K. V (x) → ∞ as x → ∞ .19 ______________________________________ Consider the nonlinear system & x1 = −6 x1 + 2 x 2 3 & x2 = 2 x1 − 6 x2 − 2 x 2 Note that the ith component ∇Vi is simply the directional derivative ∂V / ∂xi .2).0.21) Since V (x) → ∞ as x → ∞ . i. In order to recover a unique scalar function V from the gradient ∇V . the matrix F (x) = A T P + PA + Q is negative semi-definite in some neighborhood Ω of the origin. If the region Ω is the whole state space.0) dx1 + ∫ x2 0 ∇V2 ( x1 . in addition.. the gradient function has to satisfy the so-called curl conditions ∂∇Vi ∂∇V j = ∂x j ∂xi (i. because the Jcobians of many systems do not satisfy the negative definiteness requirement.K.e.7 (Generalized Krasovkii Theorem) Consider the autonomous system defined by (3. n) x →∞ . The basic idea of the method is simply to check whether this particular choice indeed leads to a Lyapunov function.0. This leads to the following procedure for seeking a Lyapunov function V • assume that ∇V is given by (3. it is difficult to check the negative definiteness of the matrix F for all x.21) (or another form) • solve for the coefficients aij so as to sastify the curl equations & • assume restrict the coefficients in (3.K.2002. ∂V / ∂x n }T . ∂f A(x ) = ∂x If the matrix F = A + A T is negative definite in a neighborhood Ω . Example 3 .K.21) so that V is negative semi-definite (at least locally) • compute V from ∇V by integration • check whether V is positive definite Since satisfaction of the curl conditions implies that the above integration result is independent of the integration path. In addition. Therefore.Applied Nonlinear Control Nguyen Tan Tien . the equilibrium state at the origin is globally asymptotically stable. V (x) → ∞ as function for this system. __________________________________________________________________________________________ where the aij ’s are coefficients to be determined.5. For instance. with the equilibrium point of interest being the origin. such that ∀x ≠ 0 . let us note that a scalar function V (x) is related to its gradient ∇V by the integral relation V ( x ) = ∇V dx 0 ∫ x where ∇V = {∂V / ∂x1 .3 The Variable Gradient Method The variable gradient method is a formal approach to constructing Lyapunov functions. i. The function V (x) = f T (x) f (x) is then a Lyapunov ∫ x1 0 ∇V1 ( x1 . Theorem 3.2).0. and if in addition. Theorem 3. it is usually convenient to obtain V by integrating along a path which is parallel to each axis in turn. with the equilibrium point of interest being the origin.3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ form (3.0) dx n ___________________________________________________________________________________________________________ 16 Chapter 3 Fundamentals of Lyapunov Theory .2). then the equilibrium point at the origin is asymptotically stable. the above simply means that ∂∇V1 ∂∇V2 = ∂x2 ∂x1 We have A= ∂ f −6 2  2 ∂ x  2 − 6 − 6 x2    4 −12  F = A + AT =  2  4 − 12 − 12 x 2   The matrix F is easily shown to be negative definite. then the equilibrium point is globally asymptotically stable.2. A Lyapunov function for this system is V (x ) = f T ( x) f ( x) If Ω is the entire state space and. Let A(x) denote the Jacobian matrix of the system..

∇V = x then V can be computed 1 1 2 2 as V ( x) = ∫ x1 0 x1 dx1 + ∫ x2 0 x2 dx 2 = 2 2 x1 + x 2 2 (3. a12 = a 21 = 0 & which leads to ∇V = x .This.20 ______________________________________ Let us use the variable gradient method top find a Lyapunov function for the nonlinear system & x1 = −2x1 2 & x2 = −2 x 2 + 2 x1 x2 Estimating convergence rates for linear system Let denote the largest eigenvalue of the matrix P by λmax (P ) . 2 a 21 = 3 x2 . and therefore.23) represents another Lyapunov function for the system.18) implies that V ≤ −γ V . means that x T Q x ≤ V (0) e −γ t . Since matrix theory shows that P ≤ λmax (P ) I and λmin (Q) I ≤ Q .4 Physically motivated Lyapunov functions 3.d.d.5.26) guarantees the exponential convergence of W to zero. implies that the state x converges to the origin with a rate of at least γ / 2 .22 ______________________________________ Consider again the system in Example 3. a 22 = 3 . (3. the smallest eigenvalue of the matrix Q by λmin (Q) .26) 2 2 2 & x2 = 4 x1 x2 + x 2 ( x1 + x2 − 2) Choose the Lyapunov function candidate V = x 2 .23) λmax (P ) ≥ λmax (P0 ) Since λmin (Q1 ) = 1 = λmin (I ) .5.5 Performance analysis Lyapunov analysis can be used to determine the convergence rates of linear and nonlinear systems. 2 If the coefficients are chosen to be a11 = 1. Therefore If the coefficients are chosen to be a11 = a 22 = 1. its & derivative is V = 2V (V − 1) .22) This is indeed p. The p. a12 = x 2 . for & nonlinear systems. the satisfaction of (3. let P0 be the solution of the Lyapunov equation corresponding to Q = I is A T P0 + P0 A = −I and let P the solution corresponding to some other choice of Q A T P + PA = −Q1 Without loss of generality. Example 3 . The V (1 − V ) solution of this equation is easily found to be ___________________________________________________________________________________________________________ 17 Chapter 3 Fundamentals of Lyapunov Theory . (3.. we obtain the p. the convergence rate estimate 2 2 2 2 2 & whose derivative is V = −2 x1 − 6 x 2 − 2 x 2 ( x1 x 2 − 3 x1 x2 ) . if W is a non-negative function. the matrix (Q1 . & We can verify that V is a locally negative definite function (noting that the quadratic terms are dominant near the origin). of P and Q implies that these scalars are all strictly positive. according to xT Q x ≥ lemma. the asymptotic stability is guaranteed. and hence the above equation implies that (P . The convergence rate estimate is largest for Q = I .I) is positive semi-definite. This together with the fact xT P x ≥ λmin (P) x(t ) 2 . __________________________________________________________________________________________ γ = λmin (Q) / λmax (P ) corresponding to Q = I the larger than (or equal to) that corresponding to Q = Q1 . and therefore.P0 ) + (P .Applied Nonlinear Control Nguyen Tan Tien . we have We assume that the gradient of the undetermined Lyapunov function has the following form ∇V1 = a11 x1 + a12 x 2 ∇V2 = a 21 x1 + a 22 x2 The curl equation is ∂∇V1 ∂∇V2 = ∂x2 ∂x1 ⇒ a12 + x 2 ∂a12 ∂a = a 21 + x1 21 ∂x 2 ∂x1 λmin (Q) T x [ λmax (P ) I ] x ≥ γ V λmax (P ) & This and (3.I ) Now since λmin (Q1 ) = 1 = λmax (I ) . A simple convergence lemma Lemma: If a real function W (t ) satisfies the inequality & W (t ) + α W (t ) ≤ 0 where α is a real number. function V ( x) = 2 x1 3 2 3 + x2 + x1 x 2 2 2 (3. The difference lies in that. V and V are not necessarily quadratic function of the states. Indeed. Subtract the above two equations yields A T (P . That is dV = −2dt .d.P0 ) is positive semi-definite.2002. Estimating convergence rates for nonlinear systems The estimation convergence rate for nonlinear systems also & involves manipulating the expression of V so as to obtain an explicit estimate of V . and therefore will not affect the value of the corresponding γ . Then W (t ) ≤ W (0) e −α t The above Lemma implies that. and their ratio λmax (P) / λmin (Q) by γ .P0 ) A = −(Q1 .3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Example 3 . we can assume that λmin (Q1 ) = 1 since rescaling Q1 will rescale P by the same factor.8 2 2 2 & x1 = x1 ( x1 + x 2 − 2) − 4 x1 x 2 3.

e. .Applied Nonlinear Control Nguyen Tan Tien . This implies that the norm x(t ) of the state vector converges to zero exponentially. where α = 1 − V (0) 1 + α e −2dt 2 If x(0) = V (0) < 1 .6 Control Design Based on Lyapunov’s Direct Method There are basically two ways of using Lyapunov’s direct method for control design. then α > 0 . with a rate of 1. x & __________________________________________________________________________________________ ___________________________________________________________________________________________________________ 18 Chapter 3 Fundamentals of Lyapunov Theory . and V (t ) < α e −2t . __________________________________________________________________________________________ 3. if the trajectory starts inside the unit circle. or “explosion”). and both have a trial and error flavor: • Hypothesize one form of control law and then finding a Lyapunov function to justify the choice • Hypothesize a Lyapunov function candidate and then finding a control law to make this candidate a real Lyapunov function Example 3 . then α < 0 .e.. if the trajectory starts outside the unit circle. if V (0) > 1 .. so that V (t ) and therefore x tend to infinity in a finite time (the system is said to exhibit finite escape time. i.23 Regulator design_______________________ Consider the problem of stabilizing the system && − x 3 + x 2 = u .3 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ V (x) = α e −2dt V ( 0) .2002. i. However.

4 The equilibrium point 0 is global stable ∀x(t 0 ) . In practice. __________________________________________________________________________________________ & • The system x = − x /(1 + t ) 2 is stable (but asymptotically stable) & • The system x = − x /(1 + t ) is asymptotically stable & • The system x = −t x is exponentially stable & Another interesting example is the system x(t ) = − The solution can be expressed as x(t ) = x(t 0 ) e Since − t not x 1 + sin x 2 x ∫t0 1+sin x2 (r ) dr Extensions of the previous stability concepts Definition 4. t ) equilibrium points x* are defined by f (x* . t 0 ) such that ∫ 1 + sin x (r ) dr ≥ t0 2 t x t − t0 .6) x (t 0 ) < r ⇒ stability for non-autonomous systems both indicate the importance effect of initial time. with uniform properties have some desirable ability to ___________________________________________________________________________________________________________ 19 Chapter 4 Advanced Stability Theory . Advanced Stability Theory The objective of this chapter is to present stability analysis for non-autonomous systems. with γ t t +T being a with b(t ) ≠ 0 .1________________________________________ The system & x=− a (t ) x 1+ x2 (4. Its solution is x(t ) = x(t 0 ) e − ∫t0 a (r )dr t .4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 4. α and λ .3) Definition 4. desirable for the systems to have a certain uniformity in its behavior regardless of when the operation starts. Since Lyapunov theory is mainly developed for the stability of nonlinear systems with respect to initial conditions.2 A first-order linear time-varying system_______ has a unique equilibrium point at the origin 0 unless A(t) is always singular. implying that the system should be able to stay at the point x* all the time. For instance has an equilibrium point at x = 0 . However. Example 4.1 The equilibrium point 0 is stable at t 0 if for any R > 0 .2) (4.3 The equilibrium point 0 is exponentially stable if there exist two positive numbers. It can be regarded as a system under external input or disturbance b(t ) . It is asymptotically stable if a (r )dr = +∞ . such that for sufficiently small x(t 0 ) .2002. For instance.1) Definition 4.Applied Nonlinear Control Nguyen Tan Tien . we can easily see that the linear time-varying system & x = A(t ) x (4. the system & x=− a (t ) x 1+ x2 + b(t ) (4.9.4) & Consider the first-order system x(t ) = −a(t ) x(t ) . such problem of forced motion analysis are more appropriately treated by other methods. the equilibrium point 0 is unstable.5) ∫ a(r )dr ≥ γ . 4.1 Concepts of Stability for Non-Autonomous Systems Equilibrium points and invariant sets For non-autonomous systems. positive constant. there exists a positive scalar r ( R. does not have an equilibrium point. the system is exponentially 2 convergent with rate 1 / 2 . __________________________________________________________________________________________ Uniformity in stability concepts The previous concepts of Lyapunov stability and asymptotic x(t ) < R ∀t ≥ t 0 (4. of the form & x = f ( x. such as those in section 4. x(t ) ≤ α x 0 e −λ (t −t0 ) ∀t ≥ t 0 x(t 0 ) < r (t 0 ) ⇒ x(t ) → 0 as Note that this equation must be satisfied ∀t ≥ t 0 . t ) ≡ 0 ∀t ≥ t 0 (4. 0 ∫ ∞ It is exponentially stable if there exists a strictly positive number T such that ∀t ≥ 0 . This The definition means that we can keep the state in ball of motivates us to consider the definitions of uniform stability arbitrarily small radius R by starting the state trajectory in a and uniform asymptotic stability. x(t ) → 0 as t → ∞ Example 4. Definition 4. Non-autonomous systems ball of sufficiently small radius r . it is usually Otherwise.2 The equilibrium point 0 is asymptotically stable at t 0 if • it is stable • ∃r (t 0 ) > 0 such that t →∞ The asymptotic stability requires that there exists an attractive region for every initial time t 0 . Thus system is stable if a (t ) ≥ 0. ∀t ≥ t 0 .

all the stability properties of an autonomous system are uniform. if r = r (R) . t ) is decrescent if it is dominated by a time-invariant p. uniform asymptotic stability always implies asymptotic stability.e. Definition 4.8) Lyapunov theorem for non-autonomous system stability The main Lyapunov stability results for non-autonomous systems can be summarized by the following theorem. 4. The solution asymptotically 1+ t converges to zero.e.e. This drawback will partially be compensates by a simple result in section 4. t ) = (1 + sin 2 t )( x1 + x2 ) 2 2 V0 (x) = x1 + x2 2 2 V1 (x) = 2( x1 + x2 ) i. R2 ) Definition 4. Intuitively.. t ) ∂t ∂x ∂t ∂x dt (4. will converges into a smaller ball B R2 after a time period T which is independent of t 0 . This system has 1+ t 1 + t0 general solution x(t ) = x(t 0 ). i.8 A scalar time-varying function V (x. can be similarly applied to non-autonomous systems. By definition. the trajectory. we extend the Lyapunov analysis results of chapter 3 to the stability of non-autonomous systems.1 (Lyapunov theorem for non-autonomous systems) Stability: If. there exits a scalar function V (x.4________________________________________ Consider time-varying positive definite functions as follows 2 2 V (x. t ) = 0 and there exits a time-variant positive definite function V0 (x) such that ∀t ≥ t 0 . By uniform convergence in terms of t 0 . ∀t ≥ t 0 x(t 0 ) < R1 ⇒ x(t ) < R2 ∀t ≥ t 0 + T ( R1 .5 called Barbalat’s lemma. i.2.7 A scalar time-varying function V (x. Globally positive definite functions can be defined similarly.2 Lyapunov Analysis of Non-Autonomous Systems In this section. such that any trajectory with initial states in B R0 converges to 0 uniformly in t 0 .d. __________________________________________________________________________________________ ⇒ V (x. a time-variant function is locally positive definite if it dominates a time-variant locally positive definite function. But the convergence is not uniform. a scalar function V (x. then Time-varying positive definite functions and decrescent the equilibrium point is asymptotically stable. t ) ≥ V0 (x) (4. t ) ≤ V1 (x) .4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ withstand disturbances. whose radius is independent of t 0 .7) Thus.2002. Example 4. V is positive definite & 2. t ). ∃T ( R1 . If the condition 2 is & strengthened by requiring that V be negative definite. 4.1 Lyapunov’sdirect method for non-autonomous systems The basic idea of the direct method. this is because a larger t 0 requires a longer time to get close to the origin. and if there exits a time-variant positive definite function V1 (x) such that ∀t ≥ 0. __________________________________________________________________________________________ Given a time-varying scalar function V (x. V (x. in a ball B R0 around the equilibrium point 0.7) In other word.Applied Nonlinear Control Nguyen Tan Tien . furthermore 3. t ) with continuous partial derivatives such that 1.. t ) ≥ V1 (x) (4. its derivative along a system trajectory is ∂V ∂V d V ∂V ∂V & = + + x= f ( x. Definition 4. t ) is said to be decrescent if V (0. R2 ) > 0 such that.5 The equilibrium point 0 is locally uniformly stable if the scalar r in Definition 4. functions ___________________________________________________________________________________________________________ 20 Chapter 4 Advanced Stability Theory . V (x. then the origin is uniformly stable. V is decrescent The concept of globally uniformly asymptotic stability can be defined be replacing the ball of attraction B R0 by the whole state space.. V is negative semi-definite then the equilibrium point 0 is stable in the sense of Lyapunov. function. Example 4. a major difference in nonautonomous systems is that the powerful La Salle’s theorems do not apply.6 The equilibrium point at the origin is locally uniformly asymptotically stable if • it is uniformly stable • there exits a ball of attraction B R0 . The converse (ña o ñe ) is generally not û à true. we mean that for all R1 and R2 satisfying 0 < R2 < R1 ≤ R0 . Uniform stability and uniform asymptotic stability: If. t ) is locally positive definite if V (0. concluding the stability of nonlinear systems using scalar Lyapunov functions. t ) dominates V0 (x) and is dominated by V1 (x) because V0 (x) ≤ V (x. Definition 4. starting from within a ball B R1 . Besides more mathematical complexity.1 can be chosen independent of t 0 . The behavior of autonomous systems is dependent of the initial time. Theorem 4.3________________________________________ x & Consider the first-order system x = − . t ) = 0 . as illustrated by the following example.

∀i. the matrix A(t ) remains bounded. Perturbed linear systems Consider a linear time-varying system of the form & x = [ A1 + A 2 (t )] x (4.18) can be rewritten as x2 = x 2 (0) e .e. the ball B R0 is replaced by the whole state space. Hence..2002.17) be stable if any time t ≥ 0 .. the system is unstable. the strengthened condition 2. ∀i. its derivative along the system trajectories.Applied Nonlinear Control Nguyen Tan Tien . because it dominates the time-invariant 2 2 p.5 Global asymptotic stability_________________ Consider the system defined by & x1 (t ) = − x1 (t ) − e −2t x2 (t ) & x2 (t ) = x1 (t ) − x2 (t ) Chose the Lyapunov function candidate 2 2 V (x.. the point 0 is globally asymptotically stable. and therefore. because it is dominated by the time-invariant p. ∀t ≥ 0. condition 3. function Furthermore. Similarly to the case of autonomous systems.20) is globally stable exponentially stable.19) This can be readily shown using the Lyapunov function V = x T x . __________________________________________________________________________________________ Since LTI systems are asymptotically stable if their eigenvalues all have negative real parts ⇒ Will the system (4. t ) is radially unbounded are all satisfied. 0 ≤ xT x = V (t ) ≤ V (0) e − λt and therefore x tends to zero exponentially. Applying the above result to the first equation. is that the time-varying system (4. so does x3 . function x1 + x 2 . Example 4. 2 2 & V (x. λi [ A(t )] ≤ −α If. then V is called Lyapunov function for the nonautonomous system.18) Sufficient smoothness conditions on the A(t ) matrix Consider the linear system (4.2. 4.d. V (x. so does x 2 .2 Lyapunov analysis of linear time-varying systems Consider linear time-varying systems of the form & x = A(t ) x (4.e. then the system is globally exponentially stable. ___________________________________________________________________________________________________________ 21 Chapter 4 Advanced Stability Theory ∫ ∞ 0 A T (t ) A(t ) dt < ∞ (i.17) therefore. and (4.20) where A1 is constant and Hurwitz and the time-varying matrix A 2 (t ) is such that A 2 (t ) → 0 as t → ∞ and This function is p.8________________________________________ Consider the system defined by This shows that 2 2 2 2 & V (x. and condition 1. the integral exists and is finite) Then the system (4. the eigenvalues of A(t ) all have negative real parts ∃α > 0. is negative semi-definite. ∫ ∞ 0 A2 (t ) dt < ∞ (i. then the equilibrium point at 0 is globally uniformly asymptotically stable. the eigenvalues of A(t ) all have negative parts ? Consider the system &  x1  − 1 e 2t   x1   x2  =  0 − 1   x2  &      (4. V (x.d. V is positive definite & and V .4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Global uniform asymptotic stability: If. x1 + x1 = x 2 (0) e .d. Example 4.17). It is also decrescent. we conclude that the system is globally exponentially stable.17) is asymptotically stable if the eigenvalues of the symmetric matrix A(t ) + A T (t ) (all of which are real) remain strictly in the left-half complex plane ∃λ > 0. and & Thus. and assume that at any time t ≥ 0 . if in a certain neighborhood of the equilibrium point. t ) is negative definite. and the condition 4. The solution −t t & of (4. the integral exists and is finite) . t ) ≤ −2 ( x1 − x1 x2 + x 2 ) = −( x1 − x2 ) 2 − x1 − x 2 5 8 & x1 = −(5 + x 2 + x3 ) x1 2 & x2 = − x 2 + 4x3 & x3 = −(2 + sin t ) x3 2 Since x3 tends to zero exponentially. t ) = −2 [ x1 − x1 x2 + x 2 (1 + e −2t )] 2 x1 2 + 2x 2 .21) Both eigenvalues of A(t ) equal to -1 at all times. It is important to notice that the result provides a sufficient condition for any asymptotic stability. λi ( A(t ) + A T (t )) ≤ −λ (4. in addition. however. ∀t ≥ 0. t ) = x1 + (1 + e −2t ) x 2 A simple result. since & & & & V = x T x + xT x = x T ( A(t ) + A T (t )) ≤ −λ xT x = −λ V so that ∀t ≥ 0.

Applied Nonlinear Control Nguyen Tan Tien .23) be satisfied.3 The linearization method for non-autonomous systems Lyapunov’s linearization method can also be developed for non-autonomous systems. Let a non-autonomous system be described by (4.1 Asymptotic properties of functions and their derivatives Before discussing Barbalat’s lemma itself.2 (Barbalat) If the differentiable function f (t ) has & a finite limit as t → ∞ .23) is satisfied.A function g (t ) is said to be uniformly continuous on [0.o. the following three facts are important to keep in mind & • f → 0 ≠> f converges & The fact that f → 0 does not imply that f (t ) has a limit as t → ∞ . then it converges to a limit. g (t ) is uniformly continuous if we can always find an η which does not depend on the specific point t1 . particularly for non-autonomous systems. (4.5. Assume that f is continuously differentiable with respect to x. ⊗ Note that: . i. t ) x =0 ∀t ≥ 0 (4. then the equilibrium point 0 of the original non-autonomous system is also uniformly asymptotically stable. ∀t ≥ 0. (x. t − t1 < η ⇒ g (t ) − g (t1 ) < R or in other words. If the linearized system is only asymptotically stable. t and t1 play a symmetric role in the definition of uniform continuity. For example. ≠> & f →0 x →0 then the system & x = A(t ) x (4.Our late results require that the uniform convergence condition (4.2002. ∃η ( R ) > 0. For example.4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 4. ∃η ( R.23) satisfied) is uniformly asymptotically stable. ∞) if ∀R > 0. 4.e. ⊗ Note that: . & • If f is lower bounded and decreasing ( f ≤ 0) .o.. and if f is uniformly continuous.The linearized time-varying system must be uniformly asymptotically stable in order to use this theorem.24) is said to be the linearization (or linear approximation) of the nonlinear non-autonomous system (4. 4. a Taylor expansion of f leads to & x = A(t ) x + f h.23). the above theorem does not relate the instability of the linearized time-varying system to that of the nonlinear system.2. since it is usually very difficult to find Lyapunov functions with a negative definite derivative.5.t .23) 4. ∀t1 ≥ 0. ∞) if ∀t1 ≥ 0. t1 ) > 0. ⊗ Note that: . (4. 4.2 Barbalat’s lemma Lemma 4.22) Therem 4.1) is unstable if one or more of the eigenvalues of A 0 has a positive real part . An important and simple result which partially remedies (kha c phuï c) this situation is é Barbalat’s lemma. let us clarify a few points concerning the asymptotic properties of functions and their derivatives. Let us denote  ∂f  A(t ) =   ∂x    x =0 (4. and if (4.1) and 0 be an equilibrium point. contrary to what happens for autonomous nonlinear systems.e. as stated in the following theorem: Therem 4. & the nonlinear system x = − x + x 2 / t leads to the linearized & system x = − x . • f converges The fact that f (t ) has a limit as t → ∞ does not imply that f& → 0 . such that η does not shrink as t → ∞.and in particular. When properly used for dynamic systems.2 If the linearized system (with condition (4.t . ∀t ≥ 0. t − t1 < η ⇒ g (t ) − g (t1 ) < R .A function g (t ) is continuous on [0. . t ) If f can be well approximated by A(t ) x for any time t . regarding t as a parameter). then & f (t ) → 0 as t → ∞ . ___________________________________________________________________________________________________________ 22 Chapter 4 Advanced Stability Theory . ∀R > 0. it may lead to the satisfactory solution of many asymptotic stability problem. lim sup f h. Some non-autonomous systems may not satisfy this condition. A(t ) = A 0 ..5 Lyapunov-Like Analysis Using Barbalat’s Lemma Asymptotic stability analysis of non-autonomous systems is generally much harder than that of autonomous systems. But in some cases A is constant. (x.The Jacobian matrix A thus obtained from a nonautonomous nonlinear system is generally time-varying.4 Existence of Lyapunov Functions The for any fixed time t (i. then the instability of the linearized system implies that of the original non-autonomous nonlinear system.3 Instability Theorems 4.. Given a differentiable function f of time t .1) around equilibrium point 0. Given a non-autonomous system satisfying condition (4.Unlike Lyapunov’s linearization method for autonomous system. we can assert its (local) stability if its linear approximation is uniformly asymptotically stable. no conclusion can be draw about the stability of the original nonlinear system.e. and Lyapunov’s linearization method cannot be used for such systems. .3 If the Jacobian matrix A(t ) is constant.23) is not satisfied for the system & x = −x + t x2 . i.

The difference n − m between the order of the denominator and that of the numerator is called the relative degree of the system. t ) is negative semi-definite & • V (x. If the transfer function of the linear subsystem is so-called positive real. t1 .4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ . __________________________________________________________________________________________ This implies that V (t ) ≤ V (0) . 4. 4. because the dynamics is non-autonomous. The above lemma then follows from Barbalat’s lemma. ___________________________________________________________________________________________________________ 23 Chapter 4 Advanced Stability Theory . such that V∞ ≤ V (x(0). we study linear systems with positive real transfer function and their properties. Note that. This can be seen from the finite different theorem: ∀t .10 A transfer function h(p) is positive real if Re[h( p )] ≥ 0 for all Re[ p ] ≥ 0 (4. let us check the uniform continuity & & & of V . t ) is uniformly continuous in time & then V (x. This shows && is bounded. it means that the rational function h( p ) maps every point in the closed RHP (i. Indeed. V is uniformly continuous. and e that V & and θ were shown above to be bounded.33) is called the positive real condition. To use Barbalat’s lemma. write the system in the standard form & x = A x + Bu y = Cx Since u is bounded and the linear system is strictly stable. Geometrically. This in turn implies from the first & equation that x is bounded.e.A simple sufficient condition for a differentiable function to be uniformly continuous is that its derivative be bound. t ) is lower bounded & • V (x. one typically uses the following immediate corollary. In this section. which looks very much like an invariant set theorem in Lyapunov analysis: Lemma 4.The derivative V must be shown to be uniformly continuous. it is often possible and useful to decompose the system into a linear subsystem and a nonlinear subsystem. But the invariant set cannot be used to conclude the convergence of e . & ∃t 2 (t ≤ t 2 ≤ t1 ) such that g (t ) − g (t1 ) = g (t 2 )(t − t1 ) . Application of Babarlat’s lemma then indicates that e → 0 as t → ∞ . Example 4. And therefore. __________________________________________________________________________________________ ⊗ Note that: Such above analysis based on Barbalat’s lemma shall be called a Lyapunov-like analysis. and therefore from the second & & equation that y = C x is bounded. Thus the system output y is uniformly continuous. t ) → 0 as t → ∞ . and w(t ) is a bounded continuous function. Consider the lower bounded function V = e2 +θ 2 Its derivative is & V = 2e [−e + θ w(t )] + 2θ [−e w(t )] = −2e 2 ≤ 0 The coefficients of the numerator and denominator polynomials are assumed to be real numbers and n ≥ m . the system is not asymptotically stable. means that h( p ) always has a positive (or zero) real part when p has positive (or zero) real part. Indeed.6 Positive Linear Systems In the analysis and design of nonlinear systems. we can always use η = R / R1 independently of t1 to verify the definition of uniform continuity. & . then it has important properties which may lead to the generation of a Lyapunov function for the whole system.12_______________________________________ Consider a strictly stable linear system whose input is bounded..1 PR and SPR transfer function Consider rational transfer function of nth-order SISO linear systems.13_______________________________________ Consider the closed-loop error dynamics of an adaptive control system for a first-order plant with unknown parameter & e = −e + θ w(t ) θ& = −e w(t ) where e and θ are the two states of the closed-loop dynamics. if R1 > 0 is an upper bound on the & function g .2002.3 (Lyapunov-Like Lemma) If a scalar function V (x. and therefore. although e converges to zero. since w is bounded by hypothesis. that e and θ are bounded.33) It is strictly positive real if h( p − ε ) is positive real for some ε > 0 Condition (4. t ) satisfies the following conditions • V (x.Applied Nonlinear Control Nguyen Tan Tien . thus the state x is bounded. because θ is only guaranteed to be bounded. Example 4. V the approaches a finite limiting value V∞ . representing tracking error and parameter error. including the imaginary axis) into the closed RHP of h( p) . There are two important differences with Lyapunov analysis: .0) (this does not require uniform continuity). Hence. Definition 4. Then the system output is uniformly continuous. in addition to being negative or zero. represented in the form h( p ) = bm p m + bm−1 p m−1 + K + b0 p n + a n−1 p n−1 + K + a0 Using Barbalat’s lemma for stability analysis To apply Barbalat’s lemma to the analysis of dynamic systems. This is & typically done by proving that V& is bounded. The derivative of V is V& = −4e (−e + θ w) .The function V can simply be a lower bounded function of x and t instead of a positive definite function.6.

the real part of h( p ) is strictly positive along the jω axis.36b) The transfer function h1 . which is summarized in the celebrated Kalman-Yakubovich (KY) lemma. Of course. which is the p+λ transfer function of a first-order system. the involved system is required to be asymptotically controllable. h( p ) is a positive real function. condition (4. In fact. any asymptotically stable matrix A .4 (Kalman-Yakubovich) Consider a controllable linear time-invariant system & x = A x + bu y = cT x The transfer function h( p ) = cT [ pI − A]−1 b (4.10 A transfer function h( p) is strictly positive real (SPR) if and only if i.11 A transfer function h( p ) is positive real if.Applied Nonlinear Control Nguyen Tan Tien . and only if.. h( p) is a strictly stable transfer function ii. Its p Obviously. and of relative degree 1) function h4 actually SPR ? We have h4 ( jω ) = jω + 1 − ω + jω + 1 2 = ( jω + 1)(−ω 2 − jω + 1) (1 − ω 2 ) 2 + ω 2 In the KY lemma..16_______________________________________ Consider the transfer function of an integrator h( p ) = value corresponding to p = σ + jω is h( p ) = 1 .2002. Corresponding to the complex variable p = σ + jω . distinct) and the associated residues are real and non-negative • Re[ h( jω )] ≥ 0 for any ω ≥ 0 such that jω is not a pole of h( p ) The Kalman-Yakubovich lemma If a transfer function of a system is SPR.5 (Meyer-Kalman-Yakubovich) Given a scalar γ ≥ 0 . can be stated as follows Lemma 4. i.36a) (4. if the transfer function H (p ) = (where the second equality is obtained by multiplying numerator and denominator by the complex conjugate of the denominator) and thus Re[h4 ( jω )] = −ω 2 +1+ ω 2 (1 − ω 2 ) 2 + ω 2 = 1 (1 − ω 2 ) 2 + ω 2 γ + c T [ pI − A]−1 b 2 ___________________________________________________________________________________________________________ 24 Chapter 4 Advanced Stability Theory .. h2 and h3 are not SPR. Thus. Is the (strictly stable. __________________________________________________________________________________________ Theorem 4.15 SPR and non-SPR functions______________ Consider the following systems h1 ( p ) = h3 ( p ) = p −1 p 2 + ap + b 1 p 2 + ap + b h2 ( p ) = h4 ( p ) = p −1 p2 − p +1 p +1 p2 + p +1 Theorem 4.4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Example 4. with λ > 0 . there is an important mathematical property associated with its state-space representation. Consider the rational function h( p ) = h( p ) = which shows that h4 is SPR (since it is also strictly stable). h2 is unstable. because h1 is non-minimum phase.14 A strictly positive real function_____________ 1 .34) The above theorem implies necessary conditions for asserting whether a given transfer function h( p) is SPR: • h( p ) is strictly stable • The Nyquist plot of h( jω ) lies entirely in the RHP. ∀ω ≥ 0 Re[h( jω )] > 0 (4.9 that h( p ) is PR but not SPR. one can easily see that h( p ) is strictly positive real. • h( p ) is a stable transfer function • The poles of h( p ) on the jω axis are simple (i. __________________________________________________________________________________________ σ + λ − jω 1 = (σ + jω ) + λ (σ + λ ) 2 + ω 2 ⊗ The basic difference between PR and SPR transfer functions is that PR transfer functions may tolerate poles on the jω axis. and h3 has relative degree larger than 1.e. and only if. Lemma 4.34) can also be checked directly on a computer. for example by choosing ε = λ / 2 in Definition 4. Re[ h( p )] ≥ 0 if σ ≥ 0 . and a symmetric positive definite matrix L . while SPR functions cannot. Equivalently. vector b and c . __________________________________________________________________________________________ σ − jω . A modified version of the KY lemma. relaxing the controllability condition. minimum-phase.9. Example 4. We σ 2 +ω 2 can easily see from Definition 4.e. the phase shift of the system in response to sinusoidal inputs is always less than 900 • h( p) has relative degree of 0 or 1 • h( p ) is strictly minimum-phase (i.35) is strictly positive real if. all its zeros are in the LHP) Example 4. there exist positive matrices P and Q such that A T P + PA = -Q Pb = c (4.e.

then there exist a scalar ε > 0 .Applied Nonlinear Control Nguyen Tan Tien .3 Positive real transfer matrices The concept of positive real transfer function can be generalized to rational positive real matrices. Definition 4.4 in two aspects.2002. Such generation is useful for the analysis and design of MIMO systems. and a symmetric positive definite matrix P such that A T P + P A = -q q T − ε L Pb = c + γ q This lemma is different from Lemma 4.11 An m × m matrix H ( p ) is call PR if • H ( p ) has elements which are analytic for Re( p ) > 0 • H ( p ) + H T ( p*) is positive semi-definite for Re( p ) > 0 where the asterisk * denote the complex conjugate transpose. • the involved system now has the output equation γ u 2 • the system is only required to be stabilizable (but not necessary controllable) y = cT x + 4.7 The Passivity Formalism 4. a vector q .6. 4.10 Existence and Unicity of Solutions ___________________________________________________________________________________________________________ 25 Chapter 4 Advanced Stability Theory .8 Absolute Stability 4.9 Establishing Boundedness of Signal 4. H ( p ) is SPR if H ( p − ε ) is PR for some ε > 0 .4 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ is SPR.

6. the ~ & equivalent input v can be chosen as v = h (t ) − α h so as to d (6. A(h) is the cross section of the tank. .5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 6. the control law & v = x d ( n) − k 0 e − k1e − K − k n −1e ( n −1) (6. 6.This differs entirely from conventional linearization (such as Jacobian linearization) in that the feedback. the & resulting dynamics is linear h = v Choosing v as ~ v = −α h (6. so that the linear control theory can be applied. The dynamics (6. rather than by linear approximations of the dynamics. u ~ This implies that h (t ) → 0 as t → ∞ .2) and (6. Now. a is the cross section of outlet pipe.1 Intuitive Concepts This section describes the basic concepts of feedback linearization intuitively. Consider the system in companion form h output flow Fig.4) (where e(t ) = x(t ) − x d (t ) is the tracking error) leads to exponentially convergent tracking.1 Feedback linearization and the canonical form Example 6. α is a strictly positive constant. Feedback linearization can be applied to a class of nonlinear system described by the so-called companion form.6) ∫ where. lead to exponentially stable dynamics x ( n ) + k n −1 x ( n −1) + K + k 0 x = 0 which implies that x(t ) → 0 . b( x) : nonlinear function of the state u : scalar control input For this system.2) (6. the actual input flow is determined by the nonlinear control law u (t ) = a 2 gh − A(h) α (h) Note that in the control law (6. using the control input of the form u = (v − f ) / b (6. The control input is the flow u into the tank and the initial value is h0. using simple examples.1. 26 ___________________________________________________________________________________________________________ .Feedback linearization technique can be view as ways of transforming original system models into equivalent models of a simpler form.4) If the desired level is a known time-varying function hd (t). For tasks involving the tracking of the desired output xd (t).5) ~ still yield h (t ) → 0 when t → ∞ . or controllability canonical form.5) a 2 gh : used provide the output flow A(h) α (h) : used to rise the fluid level according to the desired linear dynamics (6. the close loop dynamics is ~ & h +α h = 0 Chapter 6 Feedback linearization (6. 6.Applied Nonlinear Control Nguyen Tan Tien .1: Controlling the fluid level in a tank Consider the control of the level h of fluid in a tank to a specified level hd.2002.1) can be rewritten as & A(h) h = u − a 2 gh If u(t) is chosen as u (t ) = a 2 gh + A(h)v (6.7) we can cancel the nonlinearities and obtain the simple inputoutput relation (multiple-integrator form) x ( n) = v .4) ~ with h = h(t ) − hd is the level error. □ ⊗ The idea of feedback linearization is to cancel the nonlinearities and imposing the desired linear dynamics.3).1 Fluid level control in a tank The dynamic model of the tank is h  d  A(h)dh  = u (t ) − a 2 gh  dt  o  & x2  x1    &    x2   x3   =  M    M     & n   f ( x ) + b( x ) u  x     where (6. From (6. Feedback Linearization Feedback linearization is an approach to nonlinear control design. Thus.1) x : the state vector f ( x). .The central idea of the approach is to algebraically transform a nonlinear system dynamics in to a fully or partly one.8) with v being an “equivalent input” to be specified.3) (6. . the & control law v = − k 0 x − k1 x − K − k n −1 x ( n −1) with the ki chosen so that the polynomial p n + k n −1 p n −1 + K + k 0 has its roots strictly in the left-half complex plane.

m2 q 2.11a) (6.11b) Fig.10) where. one may have to use algebraic transforms to Chapter 6 Feedback linearization ___________________________________________________________________________________________________________ 27 .2: Feedback linearization of a two-link robot Consider the two-link robot as in the Fig.2 Input-State Linearization Consider the problem of design the control input u for a single-input nonlinear system of the form lc2 l1 lc1 q1.12a) (6. ⊗ When the nonlinear dynamics is not in a controllability canonical form. 6.13b) (6. . in the familiar form z = A z + b v . it is not obvious at all what controller can stabilize it in a large region.Applied Nonlinear Control Nguyen Tan Tien .9) where.Find a state transformation z = z ( x ) and an input transformation u = u( x.13b) (6. q = [q1 q 2 ]T : joint angles τ = [τ 1 τ 2 ]T : joint inputs (torques) 2 2 2 H 11 = m1l c1 + I 1 + m 2 (l1 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 2 H 12 = H 21 = m2 l1c 2 cos q 2 + m2 lc 2 + I 2 2 H 22 = m 2 l c 2 + I 2 Even though linear control design can stabilize the system in a small region around the equilibrium point (0.15a) (6.11) into & z1 = −2 z1 + z 2 & z 2 = −2 z1 cos z1 + cos z1 sin z1 + a u cos(2 z1 ) (6.0). one can use the follow control law & & & & τ 1   H 11 H 12   v1  − h q 2 − h q1 − h q 2   q1   g1   =   =  &  &  +   0 τ 2   H 21 H 22  v 2   h q1  q 2   g 2  (6.2 first put the dynamics into the controllability canonical form before using the above feedback linearization design. Example: Consider a simple second order system & x1 = −2 x1 + a x 2 + sin x1 & x 2 = − x 2 cos x1 + u cos(2 x1 ) (6. Consider the following state transformation z1 = x1 z 2 = a x 2 + sin x1 which transforms (6. m 1 I2.2 A two-link robot The dynamics of a two-link robot && & & & &  H 11 H 12   q1  − h q 2 − h q1 − h q 2   q1   g1  τ 1     &&  =  &  &  +   =   H 21 H 22  q 2   h q1 0 q 2   g 2  τ 2    (6. 6. v ) .5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Example 6. and vise versa).1.2002.12b) h = m 2 l1l c 2 sin q 2 g1 = m1l c1 g cos q1 + m 2 g[l c 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m 2 l c 2 g cos(q1 + q 2 ) Control objective: to make the joint position q1 and q 2 follows desired histories q d1 (t ) and q d 2 (t ) To achieve tracking control tasks. ~ ~ & & && v = q d − 2λq − λ 2 q v = [v1 v 2 ]T : the equivalent input ~ q = q − qd : position tracking error The new state equations also have an equilibrium point at (0. so that the nonlinear system dynamics is transformed into an equivalent linear time& invariant dynamics.14) (6.14) where v is equivalent input to be designed (equivalent in the sense that determining v amounts to determining u.15b) λ : a positive number ~ ~ ~ && & The tracking error satisfies the equation q + 2λq + λ 2 q = 0 and therefore converges to zeros exponentially. τ 2 l2 & x = f ( x. leading to a linear input-state relation & z1 = −2 z1 & z2 = v Thus.11) using the original using the new transformation control input u input v (6. which cannot be directly cancelled by the control input u.0). Now the nolinearities can be canceled by the control law of the form u= 1 (v − cos z1 sin z1 + 2 z1 cos z1 ) a cos(2 z1 ) (6.12) nonlinear dynamics dynamics (6. state the problem of the problem of transformation stabilizing the original stabilizing the new (6.u ) The technique of input-state linearization solves this problem into two steps: . 6.Use standard linear technique to design v .15) input (6.τ1 I 1. A specific difficulty is the nonlinearity in the first equation.

.15).What classes of nonlinear systems can be transformed into linear systems ? . The design of tracking controller for this doubley integrator relation is simple using linear technique. Example 6. We now obtain ___________________________________________________________________________________________________________ Chapter 6 Feedback linearization 28 . perfect tracking is achieved. leading to the internal dynamics 3 To generate a direct relationship between the output and input. except at the singularity point such that x 2 = −1 .The above controller does not guarantee the stability of internal dynamics. Using the well known linear state feedback control law v = − k1 z1 − k 2 z 2 .e.3 Input-Ouput Linearization Consider a tracking control problem with the following system & x = f ( x.18b) where v is the new input to be determined.. and choosing the new input v such as & v = &&d − k1e − k 2 e y (6.21) is canceled.1. Consider the third-order system & x1 = sin x 2 + ( x 2 + 1) x 3 5 & x 2 = x1 + x 3 y = x1 Control objective: to make y track to y d (t ) 3 & & y = x1 = x 2 + u ⇒ 3 & u = − x 2 − e (t ) + y d ( t ) (6.24) The closed-loop system under the above control law is represented in the block diagram in Fig. The tracking error of the closed-loop system is given by && & e + k 2 e + k1 e = 0 (6. & & x2 + x2 = yd − e (6. and we apply a simple linear doubleintegrator relationship between the output and the new input v.16) && = ( x 2 + 1) u + f1 ( x ) y f1 ( x ) 5 = ( x1 2 + x 3 )( x 3 + cos x 2 ) + ( x 2 + 1) x1 (6. ∀t ≥ 0 . e(t ) converge to zero exponentially.k Tz u=u (x. y d (t ) and its time derivatives are assumed to be known and bounded. if initially e(0) = e(0) = 0 .5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Now.23) & resulting in the stable closed-loop dynamics z1 = −2 z1 + z 2 & and z 2 = −2 z 2 . It is linear and controllable.v) & x =f(x.21) (6. In term of the original state. k 2 = 0 or v = −2 z 2 (6. && = v . .Applied Nonlinear Control Nguyen Tan Tien . letting e = y (t ) − y d (t ) be the tracking error.17) The original state x is given from z by u= x1 = z1 x 2 = ( z 2 − sin z1 ) / a (6.29) & x3 = 2 x1 +u y = x1 Apply the same control law to the second dynamic equation.20b) (6. k 2 are positive constant. i.29) and y d is differentiate again.27b) Control objective: to make the output y (t ) track a desired trajectory y d (t ) while keeping the whole state bounded. let us & the facts that e is guaranteed to be bound by (6.3 Input-State Linearization ⊗ To generalize the above method.30) & & let us differentiate the output y = x1 = sin x 2 + ( x 2 + 1) x 3 . ⊗ Note that: .u) x where k1 .22) Clearly. one could chose k1 = 2. which is non-autonomous and nonlinear.20c) (6.Full state measurement is necessary in implementing the control law.3.20d) yields exponential convergence of e to zero. 6. 0 v=.19a) (6. consider the new dynamics (6. otherwise.25) linearization loop pole-placement loop z z=z (x) which represents an exponentially stable error dynamics. in view of & Since y is still not directly relate to the input u .21) represents an explicit relationship between y and u .19b) (6. there are two equations: . 6. & Therefore.3: Internal dynamics Consider the nonlinear control system 3 &  x1   x 2 + u   &  =  x2   u    Fig.How to find the proper transformations for those which can ? 6. (6.u ) y = h( x ) (6. If we choose the control input to be in the form u= 1 (v − f 1 ) x2 +1 (6. this control law corresponds to the original input 1 (−2 a x 2 − 2 sin x1 − cos x1 sin x1 + 2 x1 cos x1 ) a cos(2 x1 ) (6. & e+e = 0 (6.2002. However.18a) (6. the nonlinearity in (6.20a) (6. For instance.The control law is defined anywhere. then e(t ) ≡ 0.28) (6.27a) (6.

In linear systems.2 Mathematical Tools 6. given any trajectory & y d (t ) whose derivative y d (t ) is bounded.4 Input-Output Linearization of SISO System ___________________________________________________________________________________________________________ Chapter 6 Feedback linearization 29 . we have y d (t ) − e ≤ D .27b) 3 & & the out put y = x1 ≡ 0 → y = x1 ≡ 0 → u ≡ − x 2 .In nonlinear systems. ⊗ To summarize.Applied Nonlinear Control Nguyen Tan Tien . Therefore. for the system (6. . ▲ The internal dynamics of linear systems ⇒ refer the test book ▲ The zero-dynamics Definition: The zeros-dynamics is defined to be the internal dynamics of the systems when the system output is kept at zero by the input. hence the zero-dynamics is 3 & x2 + x2 = 0 (6. 6.27). Thus we can conclude from (6. .27) 3 &  x1   x 2 + u   &  =   x2   u    y = x1 (6.45) This zero-dynamics is easily seen to be asymptotically stable 2 by using Lyapunov function V = x 2 .differentiate the output y until the input u appears. the stability of the zero-dynamics implies the global stability of the internal dynamics.27a) (6.choose u to cancel the nonlinearities and guarantee tracking convergence. ⊗ The reason for defining and studying the zero-dynamics is that we want to find a simpler way of determining the stability of the internal dynamics. (6. ⊗ Note: if the second state equation in (6.study the stability of the internal dynamics. control design based on input-output linearization can be made in three steps: .30) that & & x 2 ≤ D1 / 3 . . For instance.28) does represent a satisfactory tracking control law for the system (6. where D is positive constant. since x 2 < 0 when x 2 > D1 / 3 . .5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ & assumed to be bounded.3 Input-State Linearization of SISO Systems 6.2002. if the zero-dynamics is globally exponentially stable only local stability is guaranteed for the internal dynamics. the resulting internal dynamics is unstable. and x 2 > 0 when x 2 < − D1 / 3 .27a) is replaced by & x 2 = −u .

1) Fig.1 Sliding Surfaces Consider the SI dynamic system By definition (7. 7. s 1 1 1 L p+λ p+λ p+λ 14444444444244444444443 ~ x x ( n ) = f ( x ) + b( x ) u y=x where. indeed s ≡ 0 represents a linear differential equation whose unit solution is ~ ≡ 0 . where z1 is the output of the ( n − i − 1) th filter. 7.1.1. within a short time-constant ( n − 1) /λ . ⇒ The problem of tracking the n-dimensional vector x d can be reduced to that of keeping the scalar quantity s at zero. . one has z1 ≤ φ / λn −1−i . as measured by s 2 . we get ~ ≤ Φ/ λ−λt = ε . Condition: For the tracking task to be achievable using a finite control u . Essentially. Apply the same procedure.a.1 A Notation Simplification . s (t ) ≤ Φ n −1 x ⇒ ∀t ≥ 0. given initial condition x (7.5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 7.5) where η is a strictly positive constant.3). λ is positive constant x (7. 7.b. S (t ) Bounds on s can be directly translated into bounds on the tracking error vector ~ .2 The sliding condition ___________________________________________________________________________________________________________ Chapter 7 Sliding Control 30 . bounds (7. i. where p = ( d / dt ) is the Laplace operator. Thus it constrains trajectories to point towards the surface S (t ) . decrease along system trajectories. where  d s ( x. : scalar control input u & x= x x L x (7. L . ⊗ Given initial condition (7. 1 d 2 s ≤ −η s 2 dt (7. s ~ (i) z1 x 1 1 1 1 L L p+λ p+λ p+λ p+λ 14444 244444 14444 244444 4 3 4 3 n − i −1 blocks i blocks model imprecision on f ( x ) and b ( x ) . 7. noting that ~ (i ) ≤  Φ 1 + λ  = ( 2 λ ) i ε is  x  n −1−i  λ    λ  ~(0) ≠ 0 . s (T ) dT −λ (t −T ) f ( x ) : unknown. n = 3 → s = && + 2 λ ~ + λ2 ~ . ~ (i ) can be thought of as obtained through the sequence of Fig.1.Time-varying surface S (t ) in the state-space R (n) by the scalar equation s ( x. Furthermore.2).5) states that the squared “distance” to the surface.A so-called sliding control methodology is introduced.1. 1st-order problem of keeping the scalar s at zero can now be achieved by choosing the control law u of (7. 7.a Computing bounds on ~ (i ) x From previous results. (7.2).1. unknown bounded but known sign = (Φ / λ )(1 − e − λt ) ≤ Φ / λ ..The nonlinear system with structured or unstructured uncertainties (model imprecision) is considered. 7.4) where ε = Φ/ λ . the initial desired state must be such that x d (0) = x (0) 7. Fig. t ) = 0 .Tracking error vector ~ & x x x x ≡ x − x = ~ ~ L ~ ( n −1) d (7. t ) =  + λ    dt n −1 ~ .3) ~ & & x x x x x For example. n = 2 → s = ~ + λ ~.Applied Nonlinear Control Nguyen Tan Tien .1) such that outside of S (t ) [ ] T p λ = 1− ⇒ p+λ p+λ i . In the case that x asymptotically. 7. ~ (i ) t ≤ (2 λ ) i ε. we have ∀t ≥ 0. Assume that ~ x ( 0) = 0 .2002. The simplified.2) Fig. and therefore the scalar s represents x a true measure of tracking performance. Sliding Control In this chapter: . as illustrated in Fig. the tracking error ~ is obtained from x s through a sequence of first order low-pass filter as shown in Fig.e. the problem of tracking x ≡ x d is equivalent to that of remaining on the surfaces S (t ) for all t > 0 .a Computing bounds on ~ x Let y1 be the output of the first filter y1 = From s ≤ Φ we thus get 1 n −1blocks [ ( n −1) T ] : state vector ∫e y (t ) = Φ e ∫ 0 t 0 t −λ (t −T ) s (T ) dT .4) are obtained bounded.2. i = 0. bounded nonlinear function b ( x ) : control gain. n − 1 (7.Tracking error in the variable x ~ ≡ x−x x d . x Control objective: To get the state x to track a specific time( & varying state x d = x d x d L x dn −1) [ ] T in the presence of x Similarly.

6) that the estimation error is bounded by some known function & F = F ( x. Fro instance. assume Fig.4 & x sliding mode exponential convergence finite -time reaching phase xd (t) slope . The value of α can be obtained formally from (7.6).10) & f = − a (t ) x 2 cos 3 x : unknown bounded nonlinear function ˆ with 1 ≤ a ≤ 2 .6). The other interesting aspect of the invariant set S (t ) is that once on it. 7.3) and (7. & u eq = − f + &&d − λ ~ x x (7. despite the ˆ uncertainty on the dynamics f .5).λ x s=0 A Basic Example: Consider the second-order system && = −a(t ) x 2 cos 3x + u & x where.7) & & x x To achieve s = 0 . Let f be an estimation value of f . the system trajectories are defined by the equation of the set itself.5). i.3).5 x 2 cos 3x ⇒ F = 0.3 for n = 2 .λ x s=0 Geometrically.3 Perfect Performance – At a Price Fig.. 7. the equivalent control can be constructed as u eq = α u + + (1 − α )u − i.. namely d  & x x x s =  + λ ~ = ~ + λ ~ dt   By solving (7. This intuitive construction is summarized in Fig. & x sliding mode exponential convergence finite -time reaching phase xd (t) slope . the control is chosen as & ˆ x ˆ u → u = − f + &&d − λ ~ x (7. 7.1.5) guarantees x that if condition (7.5 s<0 f+ s=0 feq fs>0 Fig.9) ˆ & & assume that f = −1. 7. 7. we choose control law as u = − f + &&d − λ ~ . The typical system behavior implied by satisfying sliding condition (7.11) We then have & & & & && s = ~ + λ ~ = ( && − &&d ) + λ ~ = f + u − &&d + λ ~ x x x x x x x (7.5) is illustrated in Fig.5) is referred to as sliding surface.12) From (7.e.5 x 2 cos 3 x . Satisfying (7. we obtain an expression for u called the equivalent control. S (t ) verifying (7.6) ~ & & & x s = ~ + λ ~ = 0 ⇒ && = −λ ~ x x x Hence. (7.5 Filippov’s construction of the equivalent dynamics in sliding mode 7.2) is not exactly verified. u eq . the surface S (t ) will be reach in a finite time smaller that s (t = 0) / η .14) 31 ___________________________________________________________________________________________________________ . which can be interpreted as the & continuous control law that would maintain s = 0 if the dynamics were exactly known. ˆ u can be seen as our best estimate of the equivalent control.8) ˆ u = u − k sgn( s) (7.4 Chattering as a result of imperfect control switchings. as a convex combination of the values of u on both side of the surface S (t ) . 7.3 Graphical interpretation of Eqs. x (0) ≠ x d (0) . : control input u y=x : scalar output of interest (7. In order to have the system track x(t ) = x d (t ) . there is chattering as shown in Fig.5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Condition (7. In order to stratify the sliding condition (7. we add to u a term discontinuous across the surface s = 0 & && = f + u eq = &&d − λ ~ x x x Chapter 7 Sliding Control (7. 7. x ) as follows ˆ f − f ≤F (7.e. ˆ Because f is unknown and replaced by its estimation f . The system‘s behavior once on the surface is called sliding regime or sliding mode. namely (d / dt + λ )n −1 ~ = 0 .2 Filippov’s Construction of the Equivalent Dynamics The dynamics while in sliding mode can be written as & s=0 (7.5) is called sliding condition.2002.Applied Nonlinear Control Nguyen Tan Tien . of course.13) And the system dynamics while in sliding mode is.n=2 When the switching control is imperfect. we have x u eq && x x x → u = − f + && = − f + ( &&d + ~ ) (7. we define a sliding surface s = 0 according to (7. which corresponds to requiring that the system trajectories be tangent to the surface.1. for a system of the form && = f + u .

letting k = F +η 1 d 2 we get from (7.15). Note that the control discontinuity has been increased in order to account for the uncertainty on the control gain b .e. ˆ it is natural to choose our estimate b of gain b as the ˆ geometric mean of the above bounds b = bmin bmax . where f − f ≤ F . Indeed. one can then easily show that the control law ˆ ˆ u = b −1 [u − k sgn( s )] with ˆ k ≥ β ( F + η ) + ( β − 1) u (7. ˆ & .5) can be rewritten as s s ≤ −η s = −η s sgn(s ) . & s = ~+λ ~ x x ~ && + λ ~ = ( && − && ) + λ ~ & & & ⇒ s=x x x xd x & & & & & ⇒ m s = m && − m &&d + m λ ~ = −c x x + u − m &&d + m λ ~ x x x x x ˆ b (7. from (7.14) to be large enough. using (7.15) satisfies the sliding condition. Indeed. Example 7. i. push hard enough in the positive direction. Integral Control A similar result would be obtained by using integral control. s = s s = [ f − f − k sgn( s )] s = ( f − f ) s − k s . instead of (7. u = − f + && − 2 λ ~ − λ2 ~ with x x ∫ ∫ d (7. It does not for higher-order systems. Since the control law will be designed to be robust to the bounded multiplicative uncertainty (7. With s and u defined as before.9): s ≤ −η s as desired.15) formally unchanged.19) & By choosing k = k ( x.8). . From (7.16) & & m && + c x x = u x (7.2002.18).5) is verified.17) can then be written in the form where x : position of vehicle u : control input (force provided by a propeller) m : mass of the vehicle c : drag coefficient In practice. β = bmax / bmin . formally letting x ∫ ~(r )dr be the variable of interest. 2 dt So that.8) is replaced by && = f + b u x (7. the sliding mode can be interpreted that “if the error is negative.13). They may more generally be functions of any measured variables external to system (7.21) where the (possibly time-varying or state-dependent) control gain b is unknown but of known bounds 0 < bmin ≤ b ≤ bmax (7. Bound (7. we can now guarantee that (7. m and c are not known accurately. we shall call β the gain margin of our ˆ design. this in turn leads to & ˆ ˆ ˆ ˆ x k ≥ b b −1 F + η b b −1 + b b −1 − 1 f − &&d + λ ~ x and thus to (7. the control discontinuity k across the surface s = 0 increases with the extent of parametric uncertainty. the integral can be defined x within a constant. Hence we have (( f − bbˆ or −1 & ˆ ˆ ˆ f ) + (1 − b b −1 )(− &&d + λ ~ ) − b b −1k sgn( s ) s x x ≤ −η s sgn( s ) ) & ˆ ˆ ˆ ˆ b b −1k s sgn( s) ≥ ( f − b b −1 f ) + (1 − b b −1 )(− &&d + λ ~ ) s x x ( ) & ˆ ˆ ˆ ˆ x x ⇒ k ≥ (b b −1 f − f ) + (b b −1 − 1)(− &&d + λ ~ ) sgn( s) + b b −1η so that k must verify & ˆ ˆ ˆ ˆ k ≥ b b −1 f − f + (b b −1 − 1)(− &&d + λ ~ ) + b b −1η x x ˆ ˆ ˆ Since f = f + ( f − f ).16) Since the control input enters multiplicatively in the dynamics.14). 2 dt ⊗ Note that: .19) in the & expression of s leads to & ˆ ˆ ˆ ˆ & s = ( f − b b −1 f ) + (1 − b b −1 )(− &&d + λ ~ ) − b b −1k sgn( s ) x x & Condition (7. The constant can be chosen to obtain s (t = 0) = 0 regardless of x d (0) . we  0x  x 0  dt    & ˆ x ˆ obtain.To the first order system.3).8) now third-order relative to this variable.12) and 1 d 2 ˆ ˆ & (7.Applied Nonlinear Control Nguyen Tan Tien .5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ sgn = +1 where “sgn” is the sign function:  sgn = −1 if if s>0 s<0 where.e.20) (7. i.. and conversely”.18) ≤β b The estimated controller is chosen as ___________________________________________________________________________________________________________ β −1 ≤ Chapter 7 Sliding Control 32 . by letting & s = ~ + 2 λ ~ + λ2 x x & x x x ∫ ~dr − ~(0) − 2 λ ~(0) 0 t Gain Margin Assume now that (7.f and F need not depend only on x or x . and may also depend explicitly on time.1________________________________________ A simplified model of the motion of an under water vehicle can be written ( ) + η s sgn( s) system (7.3) 2 t  & d   t x x gives s =  + λ   ~dr  = ~ + 2 λ ~ + λ2 ~dr . Note that be replaced by x ∫ ~(r )dr can 0 t ∫ t ~ (r )dr . and (7.14) and (7. The 0 t (7. x) in (7. because they only describe loosely the complex hydrodynamic effects that govern the vehicle’s motion. Then..20).from (7.

The constants are chosen as λ = 20.0 1.5 2. rather than apply some prepacked formula.5 x 2 cos 3 x + 0. 7. 7.0 -0.5 3.0 -1.5 2.6.6.Switching control in place of dither 7. choose the control law u as before (7.5 3.1 . for a given problem. Φ is boundary layer thickness.0 3.Out side of B (t ) .2________________________________________ & Consider again the system (7.1 : ˆ u = u − k sat ( s / φ) & & = 1.2 Continuous Approximations of Switching Control Laws In general. 7.6. and more generally guarantees that for all trajectories starting inside B(t = 0) ∀t ≥ 0. n − 1 Note that the expression (7. η = 0.Inside of B (t ) .5 1.20). reflecting the simpler structure of parametric & & uncertainty: intuitively.22) −φ u boundary layer ˆ u φ s Fig.5 1.Switching control in place of pulse-width modulation .5 x 2 cos 3 x + &&d − 20 ~ & & x x − (0. regardless of uncertainty on m .0 1.5 4 Time (s) Time (s) Fig.23) (7. & x 4 Tracking Error 4 Control input 3 2 1 0 -1 -2 -3 -4 0 0.5 4 ε Fig. 7. and the use of artificial dither to reduce stiction effects.5 x 2 cos 3 x + &&d − 20 ~ x x & & − (0.25) where.5 1.5 0 0. .7 and with smooth control law is given in Fig.0 3. it is a good idea to quickly rederive a control law satisfying the sliding condition.1) sgn( ~ + 20 ~ ) Smooth control law with a thin boundary layer φ = 0.0 3.5 1.8. K . t ) ≤ Φ} Φ>0 (7.6.5 3. ___________________________________________________________________________________________________________ Chapter 7 Sliding Control 33 . replacing in essentially assigns a low-pass filter structure to the local the expression of u the term sgn(s ) by s / Φ .The smoothing of control discontinuity inside B (t ) .1.1.0 7.a The boundary layer 0 0.0 3.5 1.20) ˆ k ≥ β ( F + η ) + ( β − 1) u & ˆ x ≥ β ( F + η ) + ( β − 1) m ( &&d − λ ~ ) + c x x x ˆ& & Hence k can be chosen as follows & ˆ k = ( F + β η ) + m ( β − 1) ( &&d − λ ~ ) x x (7.0 1.0 2. and ε = Φ/λ n-1 is the boundary layer width. Fig. chattering must be eliminated for the controller to perform properly. __________________________________________________________________________________________ Example 7. sampling time dt = 0.a illustrates boundary layer for the case n = 2 .5 3. x and assume that the desired trajectory is x d = sin(π t / 2) .23) is “tighter” than the general form (7.5 x 2 cos 3 x + 0. 6 5 1.8 Smooth control input and tracking performance __________________________________________________________________________________________ Fig.0 2. 7.2002.4 Direct Implementations of Switching Control Laws The main direct applications of the above switching controller include the control of electric motors. In general.b illustrates this concept: ⊗ Note that: .5 -1.for instance. This can be achieved by smoothing out the control discontinuity in a thin boundary layer neighboring the switching surface B (t ) = {x.b Control interpolation in the boundary layer Given the results of section 7.1.5 2.5 2.5 0.5 0. 7.Switching control with linear observer . ~ i (t ) ≤ (2 λ ) i ε x i = 0. s(x. Switching control law: ˆ u = u − k sgn( s) & & x x = 1.7 Switched control input and tracking performance ε x Control Input 6 5 4 5 4 lay er Tracking Error (x10-3) 3 2 1 0 -1 -2 -3 -4 -5 3 2 1 0 -1 -2 -3 -4 0 0. 7.0 1. 7.5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ & ˆ ˆ x u = m ( &&d − λ ~ ) + c x x x ˆ& & and a control law satisfying the sliding condition can be derived as & ˆ ˆ x u = u − k sgn( s ) = m ( &&d − λ ~ ) + c x x − k sgn( s ) x ˆ& & where k is calculated from (7.10): && = − a (t ) x 2 cos 3 x + u .5) .5 4 Time (s) Time (s) φ bo un da ry Fig. this leads to tracking to within a guaranteed precision ε .1) sat[( ~ + 20 ~ ) / φ] x x The tracking performance with switching control law is given in Fig.0 2.0 2.001 sec. interpolate to get u .Applied Nonlinear Control Nguyen Tan Tien . u can compensate for c x x directly.

Conceptually. we must actually modify condition (7.26) s 1 ( p + λ ) n −1 definition of s ~ x Fig.Applied Nonlinear Control Nguyen Tan Tien .31) (7.29) to give s . thus eliminating chattering.The boundary layer thickness φ describes the evolution of dynamics model uncertainty with time.e. & the quantity −φ is added to control discontinuity gain k (x) . instead of simply required that (7. while tracking error ~ is merely a filtered version of s x . we can now turn the boundary layer thickness φ so that (7. and whose input are. in our smooth implementation the term k (x) sgn( s ) obtained from switched control law u is actually replaced by k (x) sat( s / φ) . “perturbations”.5 xd cos 3 x d + η + λ φ & & = 0.26). the perturbations are filtered according to (7. Since λ is breakfrequency of filter (7.3) − ∆ f (x d ) + O (ε ) 1storder filter (7.5). according to definition (7.2002.5) be satisfy outside the boundary layer. uncertainty ∆ f (x d ) .1): x ( n ) = f (x) + b(x) u . d s≥φ ⇒ ( s − φ) ≤ −η dt d ⇒ ( s − φ) ≥ η s ≤ −φ dt Thus. From (7. φ = −λ φ + (0. i.4) to rewrite (7.28) in the form ___________________________________________________________________________________________________________ Chapter 7 Sliding Control 34 .29) φ We can see from (7. with s + (− ∆ f (x) + O(ε ) ) (7. . The control law is now ( )( ) ˆ where ∆ f = f − f . Indeed.11b. 7.27) sat( y ) = y if y ≤ 1  and sat is the saturation function  sat( y ) = sgn( y ) otherwise  and can be seen graphically as in the following figure sat( y ) −1 1 y ⊗ Note that: . we consider the system trajectories inside the boundary layer.5 x 2 cos 3 x + η − 0. to & s = − k ( x) the first order.30) & The additional term φ s in (7. It is thus particularly informative to plot s (t ) . & x ˆ Accordingly. it must be chosen to be “small” with respect to high-frequency un-modeled dynamics (such as unmodeled structural modes or neglected time-delays).9 Structure of the closed-loop error dynamics Control action is a function of x and x d .27) as & φ + λ φ = k (x d ) (7. whose dynamics only depend on the desired state x d . φ(t ) .1 .η ) s 2 dt (7. control law becomes u = u − k (x) sat( s / φ) .32) (7.27) can be rewritten as k ( x) = k ( x) − k ( x d ) + λ φ (7. i.3).32) & &2 k ( x ) = 0.In order to satisfy (7.. using (7. to tune up the control law so as to achieve a trade-off between tracking precision and robustness to un-modeled dynamics.28) λ = 20 .29) that the variable s (which is a measure of the algebraic distance to the surface S (t ) ) can be view as the output of the first order filter. and can be monitored so as to well exploit the control “bandwidth” available.5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ dynamics of the variable s . where & k ( x) = k ( x) − φ (7.e. In order to maintain attractiveness of the boundary layer now that φ is allowed to vary with time. . Recognizing this filter-like structure then allows us.3________________________________________ Consider again the system described by (7. we now need to guarantee that the distance to the boundary layer always decreases. in essence. Assume that φ(0) = η / λ with η = 0. 7. They can be expressed directly in terms of the variable s as & s = − k ( x) s − ∆ f ( x) φ (7. and −φ(t ) on a single diagram as illustrated in Fig.26) reflects the fact that the boundary layer attraction condition is more stringent during & boundary layer contraction ( φ < 0 ) and less stringent during & boundary layer expansion ( φ > 0 ). we now required that s ≥φ ⇒ 1 d 2 & s ≤ (φ .The s-trajectory represents a time-varying measure of the validity of the assumptions on model uncertainty.29) choice of φ ˆ b = b = 1 . It suffices to let k (x d ) =λ φ which can be written from (7.The s-trajectory is a compact descriptor of the closedloop behavior: control activity directly depends on s .5 x d cos 3 x + η ) .5 x 2 cos 3 x + η − φ & &2 where. Since k and ∆ f are continuous in x .. Furthermore.10): && = − a (t ) x 2 cos 3 x + u . which in turn provides tracking error ~ by further x low-pass filtering. Consider again the system (7.31) and (7.Boundary layer thickness φ can be made time-varying. Example 7.29) also presents a first-order filter of bandwidth λ . Now. as long as high-frequency un-modeled dynamics are not excited. Thus chattering can be eliminated.

32) become (with β d = β (x d ) ) Tracking Error (x10-2) -5 2.5 1. s. __________________________________________________________________________________________ Desired Trajectories 3 1 0 -1 -2 -3 -4 Control Input 2 20 15 10 5 0 -5 acceleration velocity (m/s) distance (m) 0 0.5): s s ≤ −η s & ˆ ((cˆ − c) x& x& + (m − m) (&x& − λ ~ ) − k sgn(s)) s ≤ −mη s x & ˆ ˆ & & ˆ& & k s sgn( s ) ≥ ((c − c) x x + c x x + (m − m) ( && − λ ~ ) ) s + mη s x x & ˆ ˆ & & k ≥ ((c − c) x x + (m − m) ( && − λ ~ ) ) sgn( s ) + mη x x d d d & m s s ≤ − mη s Tracking Error (x10-3) 4 3 2 1 0 -1 -2 -3 -4 Control Input 3 2 1 0 -1 -2 -3 -4 0 0.12 b. 7.0 1.5 1.0 1. η = 0.5 ≤ c ≤ 1. tracking error. one can easily show that (7.31) is 35 ___________________________________________________________________________________________________________ .4________________________________________ A simplified model of the motion of an under water vehicle & & can be written (7. because varying the thickness of the boundary layer allow us to make better use of the available bandwidth. The control input. .5 3.5 4 Time (s) Time (s) a.5 2. 7.0 3.0 1.5 x 2 cos 3x + &&d − λ ~ & & x x − (0. as describe above is designed as follows: & s = ~+λ ~ x x Chapter 7 Sliding Control ⊗ Remark: .If the model or its bounds are so imprecise that F can only be chosen as a large constant. 6 5 5 4 & Condition (7.5 2.33) (7.0 2.11.5 3.The arbitrary constant η (which formally.5 3.5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ ˆ u = u − k ( x) sat ( s / φ) & & x x = 1.2.5 2.0 2. so as to fully exploit our knowledge of the structure of parametric uncertainty.0 3.5 1. The a priori bounds x on m and c are: 1 ≤ m ≤ 5 and 0.5 1. The smooth control input using time-varying boundary layer.An argument similar to that of the above discussion shows that the choice of dynamics (7.The desired trajectory x d must itself be chosen smooth enough not to excite the high frequency un-modeled dynamics.5 2.5 0 -0.0 3.5 -1. Tracking error Fig. References 15 1. Control input S-trajectories (x10-2) k (x d ) ≥ λφ βd ⇒ ⇒ & φ + λ φ = β d k (x d ) & φ+ (7.5 4 -10 0 0.34) (7.11b s-trajectories with time-varying boundary layer We see that while the maximum value of the time-varying boundary layer thickness φ is the same as that originally chosen (purposefully) as the constant value of φ in Example 7. then φ from (7.36) Example 7.11a Control input and resulting tracking performance 8 6 And the controller is & u = c x x + m ( &&d − λ ~ ) ˆ ˆ& & ˆ x x  & k ( x) = max c − c x 2 + max m − m && − λ ~ + max(m)η ˆ ˆ & xd x  φ = k ( x ) − λ φ & d  k = k ( x ) − φ &  ~+λ~ & s = x x  ˆ u = u − k sat( s / φ)  The results are given in Fig.trajectories __________________________________________________________________________________________ (7. reflects the time to reach the boundary layer starting from the outside) is chosen to be small as compared to the average value of k (x d ) .5 4 & ˆ ˆ & & k ≥ (c − c) x x + (m − m) ( &&d − λ ~ ) sgn( s ) + mη x x 0 0.5 3.5 1.5 0 1 2 3 4 5 6 s λφ k (x d ) ≤ βd λφ 2 βd = k (x d ) βd −φ Time (s) k (x) = k (x) − k (x d ) + λ φ/β d with initial condition φ(0) defined as: φ(0) = β d k (x d (0)) / λ Time (s) c.5 (m/s2 ) In the case that β ≠ 1 .0 1.0 2.0 10 5 0 -5 -10 -15 -20 -25 0 1 2 3 4 5 6 b. the tracking error is consistently better (up to 4 times better) than that in Example 7.12 5 4 35 30 25 S-trajectories (x10-2) φ 4 2 0 -2 -4 -6 -8 0 0. in the sense that it guarantees the best tracking performance given the desired control bandwidth and the extent of parameter uncertainty. 7.1 . .31) and (7.Applied Nonlinear Control Nguyen Tan Tien .5 4 -5 Time (s) Time (s) Fig. Their estimate ˆ ˆ values are m = 5 and c = 1 .5 x cos 3x + η − φ) sat[( ~ + 20 ~ ) / φ] &2 & & x x s = && − &&d + λ ~ x & & & & m s = −c x x + u − m ( &&d − λ ~ ) x x & ˆ ˆ& & ˆ x u → u = c x x + m ( &&d − λ ~ ) x ˆ u = u − k sgn( s) & ˆ& & ˆ x x = c x x + m ( &&d − λ ~ ) − k sgn( s) & & ˆ x & & & ˆ& & m s = −c x x + c x x + m ( &&d − λ ~ ) − k sgn( s ) − m ( &&d − λ ~ ) x x x & ˆ ˆ & & x x = (c − c) x x + (m − m) ( && − λ ~ ) − k sgn( s) d ⊗ Note that: .3) used to define sliding surfaces is the “best-conditioned” among linear dynamics.5 s −φ 4 Time (s) Fig. λ = 20 .The value of λ is selected based on the frequency range of un-modeled dynamics.2.0 3.0 2. and s -trajectories are plotted in Fig.2002. 7.5 .0 1.35) φ 0.0 2.0 -1.21): m && + c x x = u . .5 1.

. λ typically limited by three factors: i. Thus.5 Summary (7. i = 1. . i. The desired control bandwidth λ is the minimum of three bounds (7. Thus (7. classically λ ≤ λR ≈ 2π νR 3 (7.43) λ ≤ λS ≈ ν sampling 5 where.In the case that λ is time-varying. iii.41-43). ii.44) λn ε ≈ β d k d (7.39) is particularly useful in an incremental mode. And it is not overly surprising that system performance be very sensitive to control bandwidth. sampling rate: with a full-period processing delay.A well-designed controller should be capable of gracefully handling exceptional disturbances. 7. for instance.. Neglecting time-constants of order 1 / λ . Conversely. ν sampling is the sampling rate.39) ∑b j =1 m ij ( x) u j .19): u = b −1 [u − k sgn( s )] .e. the most effective design corresponds to matching these limitations. Ideally. margin gains in performance are critically dependent on control bandwidth λ : if large λ ’s are available. structural resonant modes: λ must be smaller than the frequency ν R of the lowest un-modeled structural resonant mode. j = 1. give system model (7. condition (7.34) can be written where T A is the largest un-modeled time-delay (for instance in the actuators). It will be discussed in next section. taking notably into account the natural damping of the structural modes.Applied Nonlinear Control Nguyen Tan Tien . L . it may account for the fact that λ R may actually vary with the task. m ˆ ˆ Consider the control law (7.33) and (7. disturbances of intensity higher than the predicted bounds which are used in the derivation of the control law.1). i. neglected time delays: along the same lines. a reasonable interpretation of this constrain is. we see that the effects of parameter uncertainty on f have been “dumped” in gain k . one gets a condition of the form 1 (7. how large λ can be chosen ? In mechanical system. and conversely large modeling efforts produce only minor absolute improvements in tracking accuracy. to evaluate the effects of model simplification on tracking performance: ∆ε ≈ ∆ ( β d k d / λ n ) (7.2002.40) [ u2 L u m ]T x ( n − 2) & L x ] : the control input vector : the state vector T In particular.33)-(7. better knowledge of f reduces k by a comparable quantity.4 Multi-Input System Consider a nonlinear multi-input system of the form x i ( ni ) = f i (x) + where u = [u 1 x = x ( n −1) 7. in certain case. . having λR ≈ λ A ≈ λS ≈ λ 7.e. given clean measurements.e.3 The Modeling/Performance Trade-Offs The balance conditions (7. Furthermore. L . so that the term k sat( s / φ) simply equals λ s / β in the boundary layer. while the augmenting gain k (x) according by the quantity u ′ ( β − 1) .5 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ constant and large. the term &x ˆ u ′ = −λ ~ should be added to the corresponding u .. m .36) have practical implications in term of design/modeling/performance tradeoffs. poor dynamic models may lead to respectable tracking performance.42) ___________________________________________________________________________________________________________ Chapter 7 Sliding Control 36 . i. we have a condition for the form λ ≤ λA ≈ 1 3TA (7.41) although in practice this bound may be modulated by engineering judgment.

e. There are two main approaches for constructing adaptive controllers: so-called model-reference adaptive control method and so-called self-tuning method. This control law leads to the exponentially & ˆ x x x convergent tracking error dynamics: u = m( &&m − 2λ ~ − λ2 ~ ) .3.1 MRAC control of unknown mass____________ Consider the control of a mass on a frictionless surface by a motor force u .3 A model-reference adaptive control system A MRAC can be schematically represented by Fig.2) (8. the numbers of poles and zeros assumed to be known.5) & and the signal quantity v is defined as v = &&m − 2λ ~ − λ2 ~ . . . x x x The closed-loop zero dynamics ~ & m s + λ m s = mv Consider Lyapunov function 37 (8. λ1 : positive constants chosen to reflect the performance specifications xm : the reference model output (ideal out put of the controlled system) r (t ) : reference position * m is known exactly.1. but their locations are not. with the plant dynamics being m && = u x Choose the following model reference &&m + λ1 x m + λ2 x m = λ2 r (t ) & x (8. settling time. Existing adaptive control designs normally required linear parametrization of the controller in order to obtain adaptation mechanisms with guaranteed stability and tracking convergence. a feedback control law containing adjustable parameters. λ1 ..3) adaptation law Fig.1 Basic Concepts in Adaptive Control .2002. but that some parameters not. The controller should have perfect Chapter 8 Adaptive Control ~ x && & & & x x x x x x ⇒ m ( ~ + λ ~ ) + λ m ( ~ + λ ~ ) = m ( &&m − 2λ ~ − λ2 ~ ) Let the combined tracking error measure be & s =~ +λ ~ x x (8.How to go about designing adaptive control system ? 8. overshoot or frequency domain characteristics. 8. The main difference from conventional control lies in the existence of this mechanism. In MRAC systems.For linear plants. the adaptation law searches for parameters such that the response of the plant under adaptive control becomes the same as that of the reference model. i.What are the basic structures of adaptive control systems ? . The adaptation mechanism is used to adjust the parameters in the control law.Applied Nonlinear Control Nguyen Tan Tien . yields & m && = m ( &&m − 2λ ~ − λ2 ~ ) x ˆ x x x ~ ) ( && − 2λ ~ − λ2 ~ ).It should reflect the performance specification in the control tasks such as rise time. Adaptive Control In this chapter: . * m is not known exactly. Substitution this control law into the plant dynamics.This ideal behavior should be achievable for the adaptive control system.1) controller plant ˆ a where. It is composed of four parts: a plant containing unknown parameters.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 8. the are the are ˆ which contains the adjustable parameter m .For nonlinear plants. although parameters are unknown. The choice of the reference model has to satisfy two requirements: . with m ≡ m − m ~ ˆ & x x = (m + m xm ~ & & x x x ~ x x x ⇒ m && + 2m λ ~ + m λ2 ~ = m ( && − 2λ ~ − λ2 ~ ) m A reference model is used to specify the ideal response of the adaptive control system to external command. The controller is usually parameterized by a number of adjustable parameters.Why we need adaptive control ? . .1. Example 8.The nonlinear system with structured or unstructured uncertainties (model imprecision) is considered.2 What is Adaptive Control ? An adaptive controller differs from an ordinary controller in that the controller parameters are variable.4) ___________________________________________________________________________________________________________ . we may use the control law & ˆ x u = m ( &&m − 2λ ~ − λ2 ~ ) x x (8. and there is a mechanism for adjusting these parameters on-line based on signals in the system. The plant is assumed to have a known structure. this implies that the structure of dynamic equations is known. and an adaptation mechanism for updating the adjustable parameters. Model-Reference Adaptive Control (MRAC) ym reference model r u y e tracking capacity in order to allow the possibility f tracking convergence. with x x x ~ = x − x representing the tracking error and λ is a strictly x m positive number. .1 Why Adaptive Control ? 8.A so-called sliding control methodology is introduced. there are some inherence constrains on the structure of reference model given the assumed structure of the plant model. 8. we can choose the following control ~ & law to achieve perfect tracking && + 2λ ~ + λ2 ~ = 0 . a reference model for compactly specifying the desired output of the control system. 8.

0 -0. estimation law. 8.8 0.0 0.choose a control law containing variable parameters the pole-placement (placing the poles of the tracking error . λ2 = 25.0 0.0 2.4. 8. To increase computational efficiency.2 Self-tuning control of unknown mass________ The design of an adaptive controller usually involves the following three steps: Consider the control of a mass of Example 8.analyze the convergence properties of the resulting control but let us now generate the estimated mass parameter using a system. the unknown parameter m is slowly x time-varying.0 0.3) for generating the control input.5 1.2002. we define P (t ) ≡ 1 t Tracking Performance Parameter Estimation 0.0 1.14).0 2. furthermore.0 0.4 0. The prediction error is simply the error ˆ in fitting the known input u using the estimated parameter m . 8. x the acceleration may be close to zero.7) For simplicity. __________________________________________________________________________________________ Fig. To do this. it is desirable to adopt a recursive formulation instead of repeatedly using (8. In Fig.3 0.5 1. λ = 6. (8.8 0.6) If the update law is chosen to satisfy & ˆ m = −γ s v The derivative of Lyapunov function becomes & V = −λ m s 2 ≤ 0 (8.5 2.0 1.5 3.3 How to Design Adaptive Controllers ? Example 8.5 1.6 -0.11) 1. 8.0 1. x(0) = xm (0) = 0 and r (t ) = 0 .0 2.0 0.2 -0.5 2. For x illustration.0 Time (s) Time (s) Fig.5 and r (t ) = sin( 4t ) .0 2.0 1. The numerical values are chosen as m = 2 .0 1.5.1 0.5 2.Applied Nonlinear Control Nguyen Tan Tien .2 0.5 2.5. x(0) = x m (0) = 0 . The convergence of s to zero implies that of the & x position tracking error ~ and the velocity tracking error ~ .5 3.5 with w = && . From (8.12) dr 2 1.10) is minimal. .0 0.4 Tracking performance and parameter estimation for an unknown mass with reference path r (t ) = 0 0. 0.5 A self-tuning controller A self-tuning controller is a controller which performs simultaneous identification of the unknown plant.4 0.9) ( ) (8. the above estimate has to be recalculated at every new time instant.5 3. (8. the simplest way of estimating m is ˆ m(t ) = u (t ) &&(t ) x (8. x(0) = x m (0) = 0.Choose an adaptation law for adjusting those parameters dynamics) control laws (8.5. choosing the estimate in such a way that the total prediction error Using Barbalat’s lemma.13) and (8. & & ˆ m(0) = 0.1.11). λ1 = 10.0 The function P (t ) is called the estimation gain. with the prediction error e defined as ˆ x e(t ) ≡ m(t ) &&(t ) − u (t ) . it is easily to show that s converges to zero. If actually.5 1.8) However this is not good method because there may be considerable noise in the measurement && .5 ∫ e (r ) dr 2 0 t (8.0 0. The resulting estimating is Tracking Performance Parameter Estimation 2.2 0.5 1.1.4 and 8. 8.11)(which can be written controller u plant y ˆ P −1m = w u dr ) leads to 0 ∫ t & ˆ m = − P (t ) w e ˆ a (8. its update can be directly obtained by using d −1 P = w2 dt Time (s) Time (s) Fig. and.6 0. ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 38 . i. the results of simulation for this example are given in Fig. the parameter estimate m is obtained by numerically integrating Eqs.6 2.5 0. 8.0 0.. A better approach is to estimate the parameter using a least-squares approach.0 0.4 -0.5 3.5 ∫ w u dr ˆ m= ∫ w dr 0 t 2 0 t (8.1).5 ∫w 0 (8. Let us still use .6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 1 1 1 ~2 m s2 + m ≥0 2 2γ Its derivative yields V= ~~ & & & V = s m s + γ −1m m ~ ~ & ˆ = s (−λ m s − m v ) − γ −1m m ~ & ˆ = −λ m s 2 − γ −1m m + γ s v (8.5 Tracking performance and parameter estimation for an unknown mass with reference path r (t ) = sin( 4t ) __________________________________________________________________________________________ ( ) (8.e.0 2.5 2. Relations between MRAC and ST methods 8.0 -0.13) Self-Tuning Controller (STC) r Then differentiation of Eq. assume that the acceleration can be measured by an accelerometer. In Fig.1 0. γ = 0. This total error minimization can potentially average out the effects of measurement noise.14) estimator ˆ In implementation.0 0.

23) and (8. a m = 4 . Thus. Furthermore.24) & which lead to the closed-loop dynamics y = − a m y + bm r which is identical to the reference model dynamics.2002.27) and (8. and v (t ) is a measurable m × 1 vector. Example 8. γ = 2 . The sgn(b p ) in (8. bm > 0 are constant parameters. we get ar = bm bp ay = a p − am bp (8. the derivative of & Lyapunov function becomes V = − a m e 2 ≤ 0 .15) ˆ ˆ Now we choose the adaptation laws for a r and a y .10 ˆ ˆ If the plant parameters were known. a y = a y . a y are variable feedback gains. comparing (8. if v (t ) is bounded.3 A first-order plant________________________ & Consider the control of the unstable plant y = y + 3 u using the previous designed adaptive controller. Choice of control law As first step in adaptive controller design. Choice of reference model Let the desired performance of the adaptive control system be specified by a first-order reference model & y m = − a m y m + bm r (t ) (8.29) where. such as a r = a r . 8. the global asymptotic convergence of the tracking error e(t ) is guaranteed by ~ ~ Barbalat’s lemma.28). the closed-loop dynamics is & ˆ ˆ y = −( a p − a y b p ) y + a r b p r (8. k is an unknown constant with know sign. Choose the Lyapunov function candidate V= 1 2 1 ~ ~2 e + b p ( a r2 + a y ) ≥ 0 2 2γ (8. Furthermore.23) With adaptation laws (8.23). then e(t ) → 0 as t → ∞ .20) & ˆ a r = − sgn(b p ) γ e r & ˆ a y = − sgn(b p ) γ e y (8. b p = 3 .21) is clear: it allows the possibility of perfect model matching.1: Consider two signals e and φ related by the following dynamic equation e(t ) = H ( p ) [ k φT (t ) v (t )] (8. the signals e . the adaptive control system is globally stable. i. y is the plant output. let us choose the control law to be γ 1 ~ ~ ~ ˆ ~ ˆ & & = −a m e + e b p sgn(b p ) (a r r − a y y ) + b p ( a r a r + a y a y ) 2 = −a m e + 2 1 γ ~ ˆ & b p a r a r + sgn(b p ) γ e r + ( ) γ ˆ ˆ u = a r (t )r (t ) + a y (t ) y (t ) (8.2 Adaptive Control of First-Order Systems Let us discuss the adaptive control of first-order plants using MRAC method. and r (t ) is bounded external reference signal. Let the tracking error be e = y − y m and the error of parameter estimation be ~ ˆ ar = ar − ar ~ ˆ ay = ay − ay (8. Consider the first-order differential equation & y = −a p y + b p u (8.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Lemma 8. bm = 4 . 8. 8.28) with γ being a positive constant representing the adaptation gain.21) and (8.26) with γ being a positive constant. a r and a y & imply the boundedness of e and therefore the uniform & continuity of V .21) Its derivative yields 1 ~ ~ ~ ~ ~ ~ & & & V = e [− a m e + b p (a r r − a y y )] + b p ( a r a r + a y a y ) where a m ≥ 0.21) & ˆ ˆ e = − a m ( y − y m ) + (a m − a p + b p a y ) y + (b p a r − bm )r ~ ~ = −a m e + b p (ar r − a y y ) The Lemma 8.. H ( p) is strictly positive real transfer function. The numerical parameters are: a p = −1 . The reason for the choice of control law (8.27) (8. Two reference signals are used: * r (t ) = 4 * r (t ) = 4 sin(3t ) ⇒ simulation results in Fig. and y (0) = y m (0) = 0 . a p and b p are constant unknown plant parameters.25) where e(t ) is a scalar output signal. ~ ~ a r and a y are bounded. and yields zero tracking error. then e(t ) and φ(t ) are globally bounded. If the vector φ(t ) varies according to & φ(t ) = − sgn(k ) γ e v (t ) (8. With this control law. because the boundedness of e .1 suggests the following adaptation laws (8.22) 1 ~ ˆ & b p a y a y + sgn(b p ) γ e y γ ( ) ˆ ˆ where a r . φ(t ) is m × 1 vector function of time. Tracking convergence analysis We analyze the system’s stability and convergence behavior using Lyapunov theory.Applied Nonlinear Control Nguyen Tan Tien .e.9 ⇒ simulation results in Fig.27-28) determines the direction of the search for the proper controller parameters. Choice of adaptation law ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 39 .16) The dynamics of tracking error can be found by subtracting (8.

Adding both side of (8.33) where the second term in (8. Choice of control law Define a signal z (t ) as follows (n z (t ) = y m ) − β n−1e ( n−1) − K − β 0 e (8.K.0 2.10 Tracking performance and parameter estimation with reference path r (t ) = 4 sin(3t ) __________________________________________________________________________________________ Fig. The tracking error e = y − y m then satisfies the closed-loop dynamics a n [e ( n) + β n−1e ( n−1) + K + β 0 e = v T (t )~ (t ) (3.9 Tracking performance and parameter estimation with reference path r (t ) = 4 5 4 2.36) and rearranging. a m = 4 .0 -0.0 4. & ˆ ˆ ˆ y = −a p y − c p f ( y ) + b p [a y y + a f f ( y ) + a r r ] ˆ ˆ ˆ = − ( a p − b p a y ) y − (c p − b p a f ) f ( y ) + b p a r r Comparing to (8. Instead of using (8.0 -0.0 0. 8.0 4. The nonlinear in these dynamics is characterized by its linear parametrization in terms of the unknown constant c .0 1.5 ˆ ar ˆ ay Tracking Performance Tracking Performance Parameter Estimation 1. 8.0 1.5 2.0 1.5 5.5 4.0 1. α n y m ( n) + α n−1 y m ( n−1) + K + α 0 y m = r (t ) with r (t ) being a bounded reference signal.5 -2.11 * r (t ) = 4 sin(3t ) ⇒ simulation results in Fig.34a) (8. bm = 4 .0 4.5 5.0 -1. and y (0) = y m (0) = 0 . This represents a pole-placement controller which places the poles at positions specified by the coefficients β i . b p = 3 .5 2.5 -2. The numerical parameters are: a p = −1 .0 1.5 0.5 -2. β n being positive constants chosen such that p n + β n−1 p n−1 + K + β 0 is a stable (Hurwitz) polynomial.0 3.12 Tracking performance and parameter estimation with reference path r (t ) = 4 sin(3t ) __________________________________________________________________________________________ Parameter convergence analysis ⇒ refer text book Extension to nonlinear plant The same method of adaptive control design can be used for the non-linear first-order plant describe by the differential equation & y = −a p y − c p f ( y) + b p u (8. Using the same procedure for the linear plant.5 1.0 0.38) with β1 . 8.0 1.0 1. 8.0 -1.0 1. we can rewrite the plant dynamics as a n [ y ( n) − z ] = u − a n z − a n−1 y ( n−1) − K − a0 y Let us choose the control law to be ˆ ˆ ˆ ˆ u = a n z + a n−1 y ( n−1) + K + a0 y = v T (t )a(t ) with v (t ) = [z (t ) ˆ ˆ a(t ) = [a n & y n−1 L y ˆ ˆ a n−1 L a1 y ]T ˆ a0 ]T (3.5 ˆ ar Tracking Performance Tracking Performance Parameter Estimation 2 1 0 -1 -2 -3 -4 -5 0 1 2 3 4 5 6 7 8 9 10 1. y .22).0 2.5 1. y ( n−1) are measurable.5 -1.4 A first-order non-linear plant_______________ & Consider the control of the unstable plant y = y + y 2 + 3 u using the previous designed nonlinear adaptive controller.5 5.5 1.5 -2.0 2. now we use the control law coefficient vector a = [a n L a1 a0 ]T is unknown.0 ˆ ar ˆ ay ˆ af Time (s) Time (s) Time (s) Time (s) Fig.5 Fig.5 0.5 3.0 0.0 0.0 2.5 0.2002.0 -1.0 -1.34c) & ˆ a r = − sgn(b p ) γ e r Example 8.5 2.5 1.5 2.5 -1.5 4.0 2.0 0 1 2 3 4 5 6 7 8 9 10 Parameter Estimation 3 3 2 1 0 -1 -2 -3 -4 -5 0 1 2 3 4 5 6 7 8 9 10 1. 8.0 0 0. but their signs are known.Applied Nonlinear Control Nguyen Tan Tien .6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 6 5 4 3 2 1 0 0.5 3. Two reference signals are used: denoting the estimated parameter vector.40) a * r (t ) = 4 ⇒ simulation results in Fig. γ = 2 .0 0. The objective of the control system is to make y closely track the response of a stable reference model & where the state components y.0 -0.0 0.5 3.12 where ~ = a − a a ˆ ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 40 .5 0.5 3 1. 8.39) (8.5 2.0 3.0 0.33) is introduced with the intention of adaptively canceling the nonlinear term.34b) (8.5 0.36) where f is any known nonlinear function.0 0.21) and define ~ ˆ a f ≡ a f − a f .5 3.K.0 0 1 2 3 4 5 6 ˆ ar ˆ af ˆ ay ˆ ay 7 8 9 10 Time (s) Time (s) Time (s) Time (s) Fig.3 Adaptive Control of Linear Systems with Full States Feedback Consider the nth-order linear system in the canonical form a n y ( n ) + a n−1 y ( n−1) + K + a0 y = u (8.32) 8.37) ˆ ˆ ˆ u = a y y + a f f ( y ) + ar r (8.5 -1.0 0.0 Parameter Estimation 4. The adaptation laws are & ˆ a y = − sgn(b p ) γ e y & ˆ a f = − sgn(b p ) γ e f a f ≡ c p / bp and (8.5 1.0 2.0 3.0 1.0 -0.5 2.5 -1.11 Tracking performance and parameter estimation with reference path r (t ) = 4 5 4 2.0 0.

1 Linear systems with relative degree one Choice of the control law To determine the appropriate control law for the adaptive controller.. the high frequency response is ω essentially determined by k p .43) (8. and k m is positive. the adaptation law & ˆ a = − Γ v bT P x & leads to V = − x T Q x ≤ 0 .46) ym (t ) e r (t ) k Wm ( p ) u0 u1 u p + bm W p ( p) α1 β1 p + β 2 p + bm Fig. ~ ) = xT P x + ~ T Γ −1~ where both Γ and P are symmetric positive constant matrix. Example 8. The objective of the design is to determine a control law.e. we will assume that nm − mm ≥ n − m . Therefore.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Choice of adaptation law Rewrite the closed-loop system (3. The relative degree r of this system is r = n − m . m − 1) and the high frequency gain k p are all assumed to be unknown.13.the relative degree n − m is known .4.e. 8. We assume as follows . cT =  M       1  0  0  0  1 − β n−1       0 theory that the relative degree of the reference model has to be larger or equal to that the plant in order to allow the possibility of perfect tracking. and P satisfies PA + A T P = −Q Q = QT > 0 & for a chosen Q .Applied Nonlinear Control Nguyen Tan Tien . the transfer kp a m1 − a p1 β1 = function from the reference input r to the plant output is Wry = k m ( p + bm ) p 2 + a m1 p + a m2 = Wm ( p ) where Z m and Rm are monic Hurwitz polynomials of degrees nm and mm .the sign of k p is known . with the control law being u = α1 z + where k p is called the high-frequency gain. i. Many controller structures can be used for this purpose.45) and the reference model ym = k m ( p + bm ) p 2 + a m1 p + a m2 r (8.4 Adaptive Control of Linear Systems with Output Feedback Consider the linear time-invariant system presented bu the transfer function W ( p) = k p Z p ( p) R p ( p) = kp b0 + b1 p + K + bm−1 p m−1 + p m a0 + a1 p + K + a n−1 p n−1 + p n (8.2002. β 2 ..47) where z = u /( p + bm ) .5 A controller for perfect tracking_____________ Consider Lyapunov function candidate a a a V (x. b =  M  . and α1 . k are controller parameters. and an associated adaptation law.K. in our treatment. so that the plant output y asymptotically approaches y m .the plant is minimum phase 8.41b) L L O L L  1 0       0  0  0  M  . 8. k= be α1 = b p − bm .44) β1 p + β 2 y+kr p + bm (8. z is the output of a first-order filter with input u .1. The following one is particularly convenient for later adaptation design. b j (i = 0. If we take kp these . j = 0. In our adaptive control problem. we must first know what control law can achieve perfect tracking when the plant parameters are perfect known. β1 . The reason for this term is that the plant frequency response at high frequency kp verifies W ( jω ) = n−m .1. The derivative V can be computed easily as & & V = − xT Q x + 2~ T v b T P x + 2 ~ Γ −1~ a a a Therefore.13 Model-reference control system for relative degree 1 Let the controller be chosen as shown in Fig.41a) (3. It is well known from linear Chapter 8 Adaptive Control ___________________________________________________________________________________________________________ 41 .K. km . The desired performance is assumed to be described by a reference model with transfer function Wm ( p ) = k m Zm Rm (8. i. the coefficients ai . n − 1.40) in state space form & x = A x + b [(1 / a n ) v T ~ ] a e = cx where 1 0  0  0 1  0 A= M M M  0 0  0 − β  0 − β1 − β 2 (3.the plant order n is known . β2 = parameters a m2 − a p2 kp to .42) Consider the plant described by y= k p ( p + bp ) p + a p1 p + a p2 2 u (8. 8.

The scalar gain k * is defined to be k * = k m / k p and is intended to modulate the high frequency gain of the control system.2002. Choice of adaptation law For the sake of simplicity. perfect tracking is achieved with this control law. With the control law (8. y = k p ( p + bm ) p + bm k p ( p + b p ) = 2 p + bp p2 + a p p + a p p + a p1 p + a p2 1 2 * .Applied Nonlinear Control Nguyen Tan Tien . with φT (t ) ω / k * regarded as an external signal.The vector θ1 contains (n − 1) parameters which intend to cancel the zeros of plant. As before.The block for generating the (n − 1) × 1 vector ω 2 has the & same dynamics but with y as input. The poles of the matrix Λ are chosen to be the same as the roots of polynomial Z m ( p ) . since the transfer function from u1 to y is Wu1.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Therefore. which can be described by & ω1 = Λ ω1 + hu . y = Wu1.the reference signal r (t ) .51) becomes u (t ) = θ T (t ) ω (t ) (8. the output of the plant is y (t ) = B( p) * u (t ) = Wm r (t ) A( p ) (8. together * with the scalar gain θ 0 can move the poles of the closed-loop control system to the locations of the reference model poles. the closed-loop system has the desired transfer function. Chapter 8 Adaptive Control ___________________________________________________________________________________________________________ 42 . define as follows θ(t ) = [k (t ) θ1 (t ) θ 2 (t ) θ 3 (t )]T ω(t ) = [r (t ) ω1 (t ) ω 2 (t ) ω 3 (t )]T Then the control law (8. The resulting structure of the control * system is shown in the Fig. note that the control input in (8.47) is composed of three parts: . the control input in this system is a linear combination of: . ym (t ) Wm ( p ) r (t ) u0 u1 e u W p ( p) k* ω1 * Λ. θ* and θ* 2 0 represents controller parameters which lead to perfect tracking when the plant parameters are known.15. we can see the reason for assuming the plant to be minimumphase: this allows the plant zeros to be cancelled by the controller poles. where k * . and h is constant vector such that ( Λ.. As a result of the above three parts.49) where. det[ pI − Λ ] = Z m ( p ) (8. i. h θ2 * θ0 since these parameters result in perfect tracking. 8. * .The block for generating the filter signal ω1 represent an (n − 1) th order dynamics.49).49) .e.14 A control system with perfect tracking The structure of this control system can be described as follows: . i.50) The above controller in Fig. as * u * (t ) = k *r + θ1 ω1 + θ* ω 2 + θ* y 2 0 . θ1 . where ω1 is an (n − 1) × 1 vector.14. At this point. 8. Why the closed-loop transfer function can become exactly the same as that of the reference model ? To know this. Therefore.e. ω 2 = Λ ω 2 + hy . The control input can be rewritten in terms of the adjustable parameters and the various signals. the control system with variable gains can be equivalently represented as shown in Fig. __________________________________________________________________________________________ Corresponding to this control law and any reference input r (t ) . then θ(t ) = θ* + φ(t ) . Instead (8. the control law is chosen to be u (t ) = k r + θ1ω1 + θ 2 ω 2 + θ 0 y (8.48) Let the ideal value of θ be θ* and the error φ(t ) = θ(t ) − θ* .52).the vector signal ω1 obtained by filtering the control input u .13 can be extended to any plant with relative degree one. y 1 + W f Wu1.The second part places the closed-loop poles at locations of those of reference model. In adaptive control problem. 8..the vector signal ω 2 obtained by filtering the plant output y and the output itself. y (t ) = y m (t ).The third part of the control law (k m / k p ) r obviously replaces k p . the control law (8. by k m . the plant parameters are unknown.52) can also be written as u (t ) = θ* ω + φT (t )ω . θ1 . 8.53) T .The vector θ 2 contains (n − 1) parameters which. θ 2 and θ 0 are controller parameters to be provided by the adaptation law. k .The first part in effect replaces the plant zero by the reference model zero. h θ1 ω2 * Λ. ∀t ≥ 0 .52) Fig. the high frequency gain of the plant. The output here must be y (t ) = Wm ( p )r + Wm ( p )[φT ω / k * ] (8. Λ is an (n − 1) × (n − 1) matrix. y = k p ( p + bm ) p + (a p1 + β1k p ) p + (a p2 + β 2 k p ) 2 (8.. i.e. and the ideal control parameters described above are also unknown. This is seen by noting that the transfer function from u 0 to y is Wu0 . h) is controllable.

where the polynomial C ( p) contains the parameter in the vector θ1 . the same control structure as given in Fig. Clearly.48). we can show that the tracking error in the above adaptive control system converges to zero asymptotically. the tracking error is seen to be related to the parameter error by the simple equation e(t ) = Wm ( p ) [φT (t ) ω(t ) / k * ] (8.6 _______________________________________ Consider the second-order plant described by the transfer function y= kp p + a p1 p + a p2 2 For a general plants of relative degree larger than 1.2002. Note that the order of the filters in the control law is still (n − 1) . Example 8. since the model numerator polynomial Z m ( p ) is of degree smaller than (n − 1) . 8. 8.57) where λ ( p ) = det[ pI − Λ ] and λ1 ( p) is a Hurwitz polynomial of degree (n − 1 − m) . Specifically.1 and through a straightforward procedure for establishing signal boundedness. and the polynomial D( p ) contains the parameter in the vector θ 2 . 8. we now choose λ ( p ) = Z m ( p ) λ1 ( p ) (8. Linear system with higher relative degree The design of adaptive controller for plants with relative degree larger than 1 is both similar to. 8.54) ( p + λ 0 + α 1 )( p + a p1 p + a p2 ) + k p ( β 1 p + β 2 ) Since this is the familiar equation seen in Lemma 8. 8. the choice of control law is quite similar but the choice of adaptation law is very different. and k are chosen such that ( p + λ 0 + α 1 )( p 2 + a p1 p + a p2 ) + k p ( β 1 p + β 2 ) = ( p + λ 0 )( p 2 + a m1 p + a m2 ) and k = k m / k p . Based on Lemma 8. Then the closed-loop transfer function is easily found to be u and the reference model ym = km p + a m1 p + a m2 2 r ym (t ) e Wry = Wm ( p ) k k p Z p λ1 ( p ) Z m ( p ) R p ( p )[λ ( p ) + C ( p )] + k p Z p D( p ) (8.Applied Nonlinear Control Nguyen Tan Tien . The closed-loop transfer function from the reference signal r to the plant output y is * θ2 Λ. Choice of control law Let us start from a simple example. there exists choice of values for k .14 is chosen. h ω2 Λ. due to the assumed positiveness of k m . With this choice. the desired zeros of the reference model can be imposed. h * θ0 Fig. θ1 .4.16. and different from.13 has been replaced by a positive number λ . 8. θ 2 and θ 0 such that the above transfer function becomes exactly the same as Wm ( p ) . __________________________________________________________________________________________ where γ is positive number representing the adaptation gain and we have used the fact that the sign of k * is the same as that of k p .55) Therefore. Instead. Let us define the transfer function of the feed-forward part u / u1 of the controller by λ ( p ) /(λ ( p ) + C ( p )) . it is no longer possible to choose the poles of the filters in the controller so that det[ pI − Λ] = Z m ( p ) as in (8. However.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ r (t ) u0 k φ ω k* T * u1 u * θ1 ω1 W p ( p) y Let the controller be chosen as shown in Fig. and that of the feedback part by D( p ) / λ ( p ) . Noting that bm in the filter in Fig. such choice of parameters exists and is unique.59) u p + λ0 W p ( p) α1 β1 p + β 2 p +λ 0 The answer to this question can be obtained from the following lemma Fig.1. β 2 .16 Model-reference control system for relative degree 2 ___________________________________________________________________________________________________________ 43 Chapter 8 Adaptive Control .2.58) r (t ) k u0 u1 The question now is whether in this general case. the following adaptation law is chosen & θ = − sgn(k p ) γ e(t ) ω(t ) (8.15 An equivalent control system for time-varying gains W ry = k 1+ = kp p + λ0 p + λ0 + α1 p 2 + a p p + a p 1 2 kp p + λ0 β1 p + β 2 p + λ0 + α 1 p + λ0 p 2 + a p p + a p 1 2 k k p ( p + λ0 ) 2 Since y m (t ) = Wm ( p ) r . then the closed-loop transfer function W ry becomes identically the same as that of the reference model. that for plants with relative degree 1. or equivalently R p (λ ( p ) + C ( p )) + k p Z p D( p ) = λ1Z p Rm ( p ) (8. if the controller parameters α 1 . β 1 .

the controller parameters θ(t ) and the parameter α (t ) for forming the augmented error are updated by sgn(k p ) γ ε ω & θ=− 1 + ωT ω & α =− (8. the controller (8.e. 8.62). we consider the case of SISO system. since the estimated parameter vector θ(t ) and the signal vector ω(t ) are both available. This lemma can be used straight forward to answer our question regarding to (8. the adaptation law (8.67a) (8.65) (5.Firstly.Secondly. t ) = u i i i =1 n (8. let us write α (t ) in the form ( ( y r n−1) = y dn−1) − λn−2 e ( n−2) − K − λ0 e . define a combined error s = e ( n−1) + λn−2 e ( n−2) + K + λ0 e = ∆ ( p ) e .62)-(8.64) where α (t ) is a time-varying parameter to be determined by adaptation. then there exist polynomials M ( p) and N ( p) such that A( p ) M ( p ) + B( p ) N ( p ) = A* ( p ) where A* ( p ) is an arbitrary polynomial. without unstable hidden modes or dynamics) by the control input if the parameters are known. Choice of adaptation law When the plant parameters are unknown.63) x= y f i ( x. η (t ) can be computed on-line. From (8. For convenience.Applied Nonlinear Control Nguyen Tan Tien .67b) (8. (8..54) e(t ) = Wm ( p ) [φ (t ) ω(t ) / k ] T * ε (t ) = where 1 k* φT (t ) ω + φα η (t ) (8. A famous technique called error augmentation can be used to avoid the difficulty in finding an adaptation law for (8.62) With the control law (8.17. Note that s ( ( can be rewritten as s = y ( n−1) − y r n−1) with y r n−1) is defined as ε (t ) = e(t ) + α (t )η (t ) (8. In this section.67).64) we obtain ___________________________________________________________________________________________________________ 44 Chapter 8 Adaptive Control . the nonlinear plant dynamics can be linearly parameterized. t ) [ & y L y ( n−1) ] T : the state vector : known nonlinear functions : unknown constant as shown in Fig.60) ω(t ) = Wm ( p )[ω] (8. respectively. Using the gradient method with normilazation. 2.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Lemma 8.68) can be rewritten in the form h y ( n) + ∑ a f ( x.66) This implies that the augmented error can be linearly parameterized by the parameter error φ(t ) and φα . It is useful to note two features about this error . 8. The basic idea of the technique is to consider a so-called augmented error ε (t ) which correlates to the parameter error φ in a more desirable way than the tracking error e(t ). Choice of control law Similarly to the sliding control approach.52) is used again u (t ) = θT (t ) ω(t ) and the tracking error from (8. the full state is measurable 3. t ) = b u i i i =1 n (8.61) γ εη 1 + ωT ω (8. global convergence of the tracking error can be shown.61) and adaptation law (8. α (t ) = 1 / k * + φα (t ) .55) cannot be used.17 The augmented error Let define an auxiliary error η (t ) by ∑ α f ( x. h = 1 / b and ai = α i / b .70) where.5 Adaptive Control of Nonlinear System Consider a class of nonlinear system satisfying the following conditions: 1. in the sense that when θ(t ) is replaced by the true (constant) parameter vector θ* .59). we have θ*Wm ( p )[ω(t )] − Wm ( p )[θ* (t ) ω(t )] = 0. η (t ) is caused by time-varying nature of the estimated parameters θ(t ) .2: Let A( p ) and B( p ) be polynomials of degree n1 and n2 . . ω (t ) φT θT W p ( p) W p ( p) θT 1 k* e(t ) ε (t ) η (t ) α (t ) W p ( p) Fig.68) η (t ) = θT (t )Wm ( p )[ω(t )] − Wm ( p )[θT (t ) ω(t )] (8.2002. where the output tracking error is e = y − y d and a stable (Hurwitz) polynomial is ∆ ( p ) = p n−1 + λn−2 p n−2 + K + λ0 . If A( p ) and B( p ) are relative prime. This also implies that η can be written: η (t ) = φT Wm (ω) − Wm (φT ω) Define an augmented error ε (t ) αi . Problem statement Consider nth-order nonlinear systems in companion form y ( n) + where However. 8.b The control objective is track a desired output y d (t ) despite the parameter uncertainty. nonlinearities can be cancelled stably (i.

some ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 45 . (8.72) ∑ aˆ f (x.low-frequency un-modeled dynamics. Because the unknown parameters.. k = − γ s f 2 x & ~ ~ ~ Lyapunov function: V = m s 2 + γ −1 ( m 2 + c 2 + k 2 ) and its derivative with the above adaptation laws yields ~ & ~& ~& ˆ & & ˆ ˆ V = 2m s s + 2γ −1 ( m m + c c + k k ) ~ ~x ~ ~ & ~& ~& ˆ ˆ ˆ = 2 s (m &&r + c f1 + k f 2 − α s) + 2γ −1 ( m m + c c + k k ) ~ ˆ & ~ ˆ ~ ˆ & & x = −2α s 2 + 2γ −1  m(m + γ s &&r ) + c (c + γ s f1 ) + k (k + γ s f 2 )      = −2α s 2 ≤ 0 __________________________________________________________________________________________ (8. in practice. t )  r  p + ( k / h)  i =1   (   y r n)      1/ h  ~ ~ ~  f1 (x. is (n obtained by modifying y d ) according to the tracking errors. Apply the above analysis procedure.2002.71) ∑ a f (x.what effects can they have on adaptive control systems ? . (n Noting that y r ) . the x x controller is ˆ ˆ ˆx ˆ u → u = m &&r − α s + c f1 + k f 2 which leads to the tracking error ˆ ˆx ˆ m && + c f1 + k f 2 = m &&r − α s + c f1 + k f 2 x ˆ ˆx ˆ (m && − m &&r ) − (m &&r − m &&r ) − (c − c) f1 − (k − k ) f 2 + α s = 0 x x x ~ ~ && − c f − k f + α s = 0 ~ m ( && − && ) − m x x x r r 1 2 the parameters are all known. If which yields m && + c f1 + k f 2 = m &&r − α s + c f1 + k f 2 or x x m ( && − &&r ) + α s = 0 .noise . c = − γ s f1 . i. such as actuator dynamics or structural vibrations .Applied Nonlinear Control Nguyen Tan Tien . The analysis has been carried out assuming that no other uncertainties exist in the control system besides parametric uncertainties. y r n) = y dn) − λn−2 e ( n−1) − K − λ0 e .71) Chose the control law u = m &&r − α s + c f1 + k f 2 x ( → 8. Example___________________________________________ Consider a mass-spring-damper system with nonlinear friction and nonlinear damping described by the equation & m && + c f1 ( x) + k f 2 ( x) = u x (8.73) can be written in the form s= n  1 / h  ~ ( n) ~ h y + ai f i (x.74) & & ˆ ˆ ˆ Adaptation laws: m = − γ s &&r . t ) i i i =1 n (n where k is constant of the same sign as h .74) 8. the control law (8.computation round-off error and sampling delay Since adaptive controllers are designed to control real physical systems and such non-parametric uncertainties are unavoidable. it is important to ask the following questions concerning the non-parametric uncertainties: .high-frequency un-modeled dynamics.73) ~ ˆ ~ ˆ where. many types of non-parametric uncertainties can be present.when are adaptive control systems sensitive to them ? . t ) i i i =1 n x  &&r  ~ ~ ~   & m s + α s = m c k  f1   f2    [ ] ( → 8.6 Robustness of Adaptive Control System The above tracking and parameter convergence analysis has provided us with considerable insight into the behavior of the adaptive control system.69) While precise answers to such questions are difficult to obtain. t )     ∑ [ ] (8.72) where h.71) is replaced by ˆ ( u = h y r n) − k s + ( → 8. the so-called “reference” value of y (n) . this choice leads to the tracking & error dynamics h s + k s = 0 and therefore gives exponential convergence of s . The tracking error yields ~ ( & h s + k s = h y r n) + ~ ∑ a f ( x. such as Coulomb friction and stiction . t )   =  h a1 L a n   p + ( k / h)  M      f n (x. These include . ai = ai − ai . Choice of adaptation law For our adaptive control. t ) i i i =1 n (8.how can adaptive control systems be made insensitive to them ? Lemma 8. and y r ) is the ( ( ( & derivative of y r n−1) . using the Lyapunov function candidate V = h s +γ 2 −1  2    h + ∑ i =1 n ai2  ≥ 0    & it is straight forward to verify that V = −2 k s 2 ≤ 0 and therefore the global tracking convergence of the adaptive control system can be easily shown.e.1 suggests the following adaptation law & ( ˆ h = − γ sgn(h) s y r n) & ˆ a = − γ sgn(h) s f i i Specially. ai have been replaced by their estimated values. hi = hi − hi . However. which in turn.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Consider the control law ( u = h y r n) − k s + & & & & & s ≡ e + λ0 e = x − ( x d − λ0 e) ≡ x − x r (8.measurement. guarantees the convergence of e . Define the error because adaptive control systems are nonlinear systems.

it is assumed that there is some measurement noise m(t ) in the adaptive system.18 Adaptive control with un-modeled dynamics and measurement noise Corresponding to the reference input r (t ) = 2 . this leads to the form y (t ) = y f (λ f − a1 ) + u f b1 (8. which are highfrequency but lightly-damped poles at (−15 ± j ) . The whole adaptive control system is shown in Fig. although simple.5 sin(16. Fig. let us filter (multiply) both side of the equation by 1 /( p + λ f ) (where p is the Laplace operator and λ f is a known positive constant). Beside the un-modeled dynamics. 8.79) by a constant number. Note that. The un-modeled dynamics are thus to seen to be 229 /( p 2 + 30 p + 229) .79) are the parameters (λ f − a1 ) and b1 .. the only unknown quantities in (8. i. To eliminate y in the above equation. The following example illustrates this situation.2002.7 On-line Parameter Estimation Few basic methods of on-line estimation are studied. Rearranging.1 Linear parameterization model The essence of parameter estimation is to extract parameter information from available data concerning the system. The measurement noise is assumed to be n(t ) = 0.Applied Nonlinear Control Nguyen Tan Tien . we may multiply both side of (8. gain of 1 / λ f .7. Example 8. The quite general model for parameter estimation applications is in the linear parameterization form y (t ) = W (t ) a where y ∈ Rn a∈R m (8.19.1t ) .77) The reference model has the following SPR function km 3 M ( p) = = p + am p+3 The real plant. 8. is actually quite general. In view of the global tracking convergence proven in the absence of non-parametric uncertainties and the small amount ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 46 . λ f . as seen in the following example. 8. the magnitudes of y f and u f are smaller than those of y and u by a factor of λ f at low frequencies. because the derivative of y appears in the above equation (noting that numerically differentiating y is usually undesirable because of noise & consideration). Example 8. Parameter drift When the signal v is persistently exciting.77) is simply a linear equation in terms of the unknown a. The above filtering introduces a d.e. However. i.e.78) This means that the real plant is of third order while the nominal plant is of only first order. the observed instability can seem quite surprising.c.18. 8. and finally diverges to infinity. 8.77). is assumed to have the transfer function relation y= 2 229 u p + 1 p 2 + 30 p + 229 : known “output” of the system : unknown parameters to be estimated W (t ) ∈ R n×m : known signal matrix (8.7 Rohrs’s example_________________________ Consider the plant described by the following nominal model H 0 ( p) = kp p+ap of non-parametric uncertainties present in the above example. then operates with a small oscillatory error related to the measurement noise. Since smaller signals may lead to slower estimation. Dead-zone 8. as a result of the filtering operation. The above model cannot be directly used for estimation. Continuous-time formulation is used. Any linear system can be rewritten in this form after filtering both side of the system dynamics equation through an exponentially stable filter of proper order. and that the output y and the input u are available. both simulations and analysis indicate that the adaptive control systems have some robustness with respect to non-parametric uncertainties.78) nominal y1 Fig. when the signals are not persistently exciting.9 Filtering linear dynamics__________________ Consider the first-order dynamics & y = − a1 y + b1u (8. It is seen that the output y (t ) initially converges to the vicinity of y = 2 . reference model ym (t ) 3 p+3 n(t ) ˆ ar u r (t ) e(t ) 2 229 p + 1 p 2 + 30 p + 229 un-modeled ˆ ay Assume that a1 . Model (8..6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ qualitative answers can improve our understanding of adaptive control system behavior in practical applications.b1 in model are unknown. however. even small uncertainties may lead to severe problems for adaptive controllers.19 Instability and parameter drift _________________________________________________________________________________________ where y f = y /( p + λ f ) and u f = u /( p + λ f ) with subscript f denoting filtered quantities. the results of adaptive control system are shown in Fig.

5 Least-squares with exponential forgetting ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 47 . yields & & ∫ w(t − r )τ(r )dr = ∫ w(t − r )[Hq& + Cq + G]dr 0 0 t t θ = [α 0 − a0 α1 − a1 L α n−1 − a n−1 b0 L bn−1 ] T T  y py p n−1 y u p n−1u  w= L L  A0 A0 A0 A0 A0     Note that w can be computed on-line based on the available values of y and u .81) can be rewritten in the form y = θT w (t ) where This linear parametrization property actually applies to any mechanical system. Specially.3 The gradient estimator 8. m2 q 2.83) has known coefficients. we can use the above filtering technique.τ1 l2 lc2 l1 lc1 ∫ t 0 && & w(t − r )Hq dr = w(t − r )Hq 0 − t t & ∫ dr [wH]q dr 0 t d & & = w(0)H (q)q − w(0)H[q (0)] q(0) − && & & ∫ [w(t − r )Hq . q = [q1 q 2 ]T : joint angles 2 2 2 H 11 = m1l c1 + I 1 + m 2 (l1 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 where y is the filtered torque and W is the filtered version of Y1 . In view of the fact that A0 ( p ) − A( p ) = (α 0 − a0 ) + (α1 − a1 ) p + K + (α n−1 − a n−1 ) p n−1 (8.84) Example 8. Specially. (8. The filtered torque y can also be computed because the torque signals issued by the computer are known. and Then each term on the left-hand side of (6. let w(t ) be the impulse response of a stable. its dynamics can be described by A( p ) y = B ( p ) u where A( p ) = a0 + a1 p + K + a n−1 p n−1 + p n B ( p ) = b0 + b1 p + K + a n−1 p n−1 Divide both sides of (8.7. _________________________________________________________________________________________ 8. 2 I 2 + m2 l c 2 .7.48) can be rewritten as & y (t ) = W(q.7. m 1 q1. _________________________________________________________________________________________ (8.4 The standard least-squares estimator τ = [τ 1 τ 2 ]T : joint inputs (torques) 8.85) Fig. Relation (8.81) n +p 2 H11 = a3 + a 4 + a1l1 + 2a 2 l1 cos q 2 H 22 = a 4 H12 = H 21 = a 2 l1 cos q 2 + a 4 Thus we can write & && τ = Y1 (q. a 2 = m2lc 2 . leading to A0 ( p ) − A( p) B( p ) u A0 ( p ) A0 ( p ) where y= A0 = α 0 + α1 p + K + α n−1 p n −1 2 H 12 = H 21 = m 2 l1c 2 cos q 2 + m 2 l c 2 + I 2 2 H 22 = m 2 l c 2 + I 2 (8. To avoid this.83) cannot be directly used for parameter estimation.80) h = m 2 l1l c 2 sin q 2 g1 = m1l c1 g cos q1 + m 2 g[l c 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m 2 l c 2 g cos(q1 + q 2 ) 2 Let us define a1 = m2 .Applied Nonlinear Control Nguyen Tan Tien . a3 = I1 + m1lc1 . Then. q) a (8.2 Prediction-error-based estimation model 8. τ 2 I 1.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ Generally. Thus the matrix W can be computed from available & measurements of q and q . including multiple-link robots. q)a (8. for a linear SISO system.7.9) a4 = is linear terms of the equivalent inertia parameters a = [a1 a 2 a3 a 4 ]T . proper filter. q. the first term on the right hand side of (8.9).2002.9 Linear parametrization of robot dynamics_____ Using partial integration.80) by a known monic polynomial of order n.9) where.84) can be rewritten as I2. 6.82) (8.2 A two-link robot Consider the nonlinear dynamics of a two-link robot && & & & &  H 11 H 12   q1  − h q 2 − h q1 − h q 2   q1   g1  τ 1     &&  =  &  &  +   =   0  H 21 H 22  q 2   h q1  q 2   g 2  τ 2  (6. convolving both sides of (6.w(t − r )Hq] dr 0 This means that the equation (8. because of the present of the un-measurable joint && acceleration q .

8 Composite Adaptation ___________________________________________________________________________________________________________ Chapter 8 Adaptive Control 48 .7.7 Concluding remarks and implementation issues 8.Applied Nonlinear Control Nguyen Tan Tien .6 Bounded gain forgetting 8.6 _________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________ 8.7.2002.

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.