Nonlinear Control Lecture 4: Stability Analysis I

Transcription

1 Nonlinear Control Lecture 4: Stability Analysis I Farzaneh Abdollahi Department of Electrical Engineering Amirkabir University of Technology Fall 2010 Farzaneh Abdollahi Nonlinear Control Lecture 4 1/70

2 Autonomous Systems Lyapunov Stability Variable Gradient Method Region of Attraction Invariance Principle Linear System and Linearization Lyapunov and Lasalle Theorem Application Example: Robot Manipulator Control Design Based on Lyapunov s Direct Method Estimating Region of Attraction Farzaneh Abdollahi Nonlinear Control Lecture 4 2/70

3 Stability Stability theory is divided into three parts: 1. Stability of equilibrium points 2. Stability of periodic orbits 3. Input/output stability An equilibrium point (Equ. pt.) is: Stable if all solutions starting at nearby points stay nearby. Asymptotically Stable if all solutions starting at nearby points not only stay nearby, but also tend to the Equ. pt. as time approaches infinity. Exponentially Stable, if the rate of converging to the Equ. pt. is exponentially. Lyapunov stability theorems give sufficient conditions for stability, asymptotic stability, and so on. Lyapunov stability analysis can be used to show boundedness of the solution even when the system has no equilibrium points. The theorems provide necessary conditions for stability are so-called converse theorems. Farzaneh Abdollahi Nonlinear Control Lecture 4 3/70

4 The most popular method for studying stability of nonlinear systems is introduced by a Russian mathematician named Alexander Mikhailovich Lyapunov Lyapunov s work The General Problem of Motion Stability published in 1892 includes two methods: Linearization Method: studies nonlinear local stability around an Equ. point from stability properties of its linear approximation Direct Method: not restricted to local motion. Stability of nonlinear system is studied by proposing a scalar energy-like function for the system and examining its time variation His work was then introduced by other scientists like Poincare and Lasalle Farzaneh Abdollahi Nonlinear Control Lecture 4 4/70

5 Autonomous Systems Consider the autonomous system: ẋ = f (x) where f : D R n is a locally Lip. function on a domain D R n. Let x D be an Equ. point, that is f ( x) = 0. Objective: To characterize stability of x. without loss of generality (wlog), let x = 0 If x 0, introduce a coordinate transformation: y = x x, then ẏ = ẋ = f (y + x) = g(y) with g(0) = 0 Farzaneh Abdollahi Nonlinear Control Lecture 4 5/70

6 The Equ. point x = 0 of ẋ = f (x) is: stable, if for each ɛ > 0, δ = δ(ɛ) > 0 s.t. x(0) < δ = x(t) < ɛ t 0 unstable, if it is not stable asymptotically stable, if it is stable and δ can be chosen s.t. x(0) < δ = lim x(t) = 0 t Lyapunov stability means that the system trajectory can be kept arbitrary close to the origin by starting sufficiently close to it. An Equ. point which is Lypunov stabile but not asymptotically stable is called Marginally stable Farzaneh Abdollahi Nonlinear Control Lecture 4 6/70

7 Example: Van Der Pol Oscillator Van der pol oscillator dynamics: ẋ 1 = x 2 ẋ 2 = x 1 + (1 x 1 ) 2 x 2 All system trajectories start except from origin, asymptotically approaches a limit cycle. Even the system states remain around the Equ. point in a certain sense, they can not stay arbitrarily close to it. So the Equ. point is unstable. Implicit in Lyapunov stability condition is that the sol. are defined t 0. This is not guaranteed by local Lip. The additional condition imposed by Lyapunov theorem will ensure global existence of sol. Farzaneh Abdollahi Nonlinear Control Lecture 4 7/70

8 Lyapunov Stability Physical Motivation Consider the pendulum example (recall Lecture 2): ẋ 1 = x 2 ẋ 2 = g sinx 1 k l m x 2 In first period it has two Equ. pts. (x 1 = 0, x 2 = 0) & (x 1 = π, x 2 = 0) For frictionless pendulum, i.e. k = 0 : trajectories are closed orbits in neighborhood of 1 st Equ. pt. ɛ δ requirement for stability is satisfied. However, it is not asymptotically stable. For Pendulum with friction, i.e. k > 0 the 1 st Equ. pt. is a stable focus ɛ δ requirement for asymptotic stability is satisfied. the 2 nd Equ. pt. is a saddle point ɛ δ requirement is not satisfied it is unstable Farzaneh Abdollahi Nonlinear Control Lecture 4 8/70

9 To generalize the phase-plane analysis, consider the energy associated with the pendulum: E(x) = 1 2 x x1 0 g l sin y dy = 1 2 x g l (1 cosx 1), E(0) = 0 If k = 0, system is conservative, i.e. there is no dissipation of energy: E = constant during the motion of the system. de dt = 0 along the traj. of the system. If k > 0, energy is being dissipated de dt < 0 along the traj. of the system. E starts to decrease until it eventually reaches zero, at that pt. x = 0. Lyapunov showed that certain other function can be used instead of energy function to determine stability of an Equ. pt. Farzaneh Abdollahi Nonlinear Control Lecture 4 9/70

10 Lyapunov s Direct Method: Let x = 0 be an Equ. pt. for ẋ = f (x). Let V : D R, D R n be a continuously differentiable function on a neighborhood D of x = 0, s.t. 1. V (0) = 0 2. V (x) > 0 in D {0} 3. V (x) 0 in D Then x = 0 is stable. Moreover, if V (x)<0 in D {0} then x = 0 is asymptotically stable. The continuously differentiable function V (x) is called a Lyapunov function. The surface V (x) = c, for some c > 0 is called a Lyapunov surface or level surface. Farzaneh Abdollahi Nonlinear Control Lecture 4 10/70

11 Lyapunov Stability when V 0 when a trajectory crosses a Lyapunov surface V (x) = c, it moves inside the set Ω c = {x R n V (x) c} and traps inside Ω c. when V < 0 trajectories move from one level surface to an inner level with smaller c till V (x) = c shrinks to zero as time goes on Farzaneh Abdollahi Nonlinear Control Lecture 4 11/70

12 Lyapunov Stability A function satisfying V (0) = 0 & V (x) > 0 in D {0} is said to be Positive Definite (p.d.) If it satisfies a weaker condition V (x) 0 for x 0 is said to be Positive Semi-Definite (p.s.d.) A function is Negative Definite (n.d.) or Negative Semi-Definite (n.s.d.) if V (x) is p.d. or p.s.d., respectively. Lyapunov theorem states that: The origin is stable if there is a continuously differentiable, p.d. function V (x) s.t. V (x) is n.s.d., and is asymptotically if V (x) is n.d. Note that when x is a vector: V (x) = V x f (x) = n i=1 V x i ẋ i = n i=1 V [ f i = V V x x 1 i x n ] f 1. Farzaneh Abdollahi Nonlinear Control Lecture 4 12/70 f n

13 Lyapunov Stability A class of scalar functions for which sign definition can be easily checked is quadratic functions: V (x) = x T Px = n n x i x j P ij i=1 j=1 where P = P T is a real matrix. V (x) is p.d./p.s.d. iff λ i {P} > 0 or λ i {P} 0, i = 1...n λ i {P} > 0 or λ i {P} 0, i = 1...n iff all leading principle minors of P are positive or non-negative, respectively. If V (x) is p.d. (p.s.d.), we say the matrix P is p.d. (p.s.d.) and write P > 0 (P 0). Farzaneh Abdollahi Nonlinear Control Lecture 4 13/70

14 Example 1 V (x) = ax x 1 x 3 + ax x 2 x 3 + ax3 2 = [ ] a 0 1 x 1 x 1 x 2 x 3 0 a 2 x 2 = x T Px 1 2 a x 3 The leading principle minors are [ ] a 0 1 a 0 det(a) = a; det = a 2 ; det 0 a 2 = a(a 2 5) 0 a 1 2 a V (x) is p.s.d. if a 5, V (x) is p.d. if a > 5 For n.d. the leading principle minors of P should be positive. OR P should alternate in sign with the first one neg. (odds: neg., even: pos.) V (x) is n.s.d. if a 5, V (x) is n.d. if a < 5, V (x) is sign indefinite for 5 < a < 5 Farzaneh Abdollahi Nonlinear Control Lecture 4 14/70

15 Example 2 Consider ẋ = g(x) where g(x) is locally Lip. on ( a, a) & g(0) = 0, xg(x) > 0, x 0, x ( a, a). stability? origin is Equ. pt. Solution 1: starting on either side of the origin will have to move toward the origin due to the sign of ẋ Origin is an isolated Equ. pt. and is asymptotically stable. Solution 2: using Lypunov theorem: x Consider the function V (x) = g(y)dy over D = ( a, a). 0 V (x) is continuously differentiable, V (0) = 0 and V (x) > 0, x 0. V is a valid Lyapunov candidate To see if it is really a Lyap. fcn, we have to take its derivative along system trajectory: V (x) = V x ( g(x)) = g 2 (x) < 0, x D {0} V (x) is a valid Lyap. fcn the origin is asymptotically stable. Farzaneh Abdollahi Nonlinear Control Lecture 4 15/70

16 Example 3: Frictionless Pendulum ẋ 1 = x 2 ẋ 2 = g sinx 1 l Study stability of the Equ. pt. at the origin. A natural Lyap. fcn is the energy fcn: V (x) = g l (1 cosx 1 ) x 2 2 V (0) = 0 and V (x) is p.d. over the domain 2π x 1 2π. V = g l ẋ 1 sinx 1 + x 2 ẋ 2 = g l x 2sinx 1 g l x 2sinx 1 = 0 V (x) satisfies the condition of the Lyap. Theorem origin is stable Farzaneh Abdollahi Nonlinear Control Lecture 4 16/70

17 Example 4: Pendulum with Friction ẋ 1 = x 2 ẋ 2 = g sinx 1 k l m x 2 Take the energy fcn as a Lyap. fcn candidate V (x) = g l (1 cosx 1 ) x 2 2 V = k m x 2 2 V (x) is n.s.d. It is not since n.d. since V = 0 for x2 = 0 and all x 1 0. the origin is only stable. But, phase portrait showed asymptotic stability!! Toward this end, let s choose: V (x) = 1 2 x T Px + g l (1 cosx 1 ) Farzaneh Abdollahi Nonlinear Control Lecture 4 17/70

18 [ ] P11 P where P = 12 is p.d. (P P 12 P 11 > 0, P 22 > 0, P 11 P 22 P12 2 > 0) 22 V = 1 2 (ẋ T Px + x T Pẋ) + g ẋ 1 sinx 1 = g l l (1 P 22)x 2 sinx 1 g l P k 12x 1 sinx 1 + (P 11 P 12 m )x k 1x 2 + (P 12 P 22 m )x 2 2 Select P s.t. V is n.d. (cancel sign indefinite factors: x 2 sinx 1 and x 1 x 2 ) P 22 = 1, P 11 = k m P 12, 0 < P 12 < k m P 12 = 1 2 k m ) V = 1 g 2 l (for V (x) to be p.d., take k m x 1sinx 1 1 k 2 m x 2 2 x 1 sinx 1 > 0 0 < x 1 < π, defining a domain D by D = {x R 2 x 1 < π} V (x) is p.d. and V is n.d. over D. Thus, origin is asymptotically stable (a.s.) by the theorem. Farzaneh Abdollahi Nonlinear Control Lecture 4 18/70

19 How Search for A Lyapunov Function? Lyapunov theorem is only sufficient. Failure of a Lyap. fcn candidate to satisfy the theorem does not mean the Equ. pt. is unstable. Variable Gradient Method Idea is working backward: Investigated an expression for V (x) and go back to choose the parameters of V (x) so as to make V (x) n.d. (n.s.d) Let V = V (x) and g(x) = x V = ( V x Then V = V x f = g T f Choose g(x) s.t. it would be the gradient of a p.d. fcn V and make V n.d. (n.s.d) g(x) is the gradient of a scalar fcn iff the Jacobian matrix g x ) T is symmetric: g i x j = g j x i, i, j = 1,..., n Farzaneh Abdollahi Nonlinear Control Lecture 4 19/70

20 Variable Gradient Method Select g(x) s.t. g T (x)f (x) is n.d. Then, V (x) is computed from the integral: x x n V (x) = g(y)dy = g i (y)dy i 0 The integration is taken over any path joining the origin to x. This can be done along the axes: V (x) = x g 1 (y 1, 0,..., 0)dy 1 + xn 0 0 i=1 x2 g n (x 1, x 2,..., y n )dy n 0 g 2 (x 1, y 2, 0,..., 0)dy 2 By leaving some parameters of g undetermined, one would try to choose them so that V is p.d. Farzaneh Abdollahi Nonlinear Control Lecture 4 20/70

21 Example 5: ẋ 1 = x 2 ẋ 2 = h(x 1 ) ax 2 where a > 0, h(.) is locally Lip., h(0) = 0, yh(y) > 0, y 0, y ( b, c), b, c > 0. The pendulum is a special case of this system. Find proper Lypunov function? Applying variable gradient method, we must find g(x) s.t. g 1 x 2 = g 2 x 1 V (x) = g 1 (x)x 2 g 2 (x)(h(x 1 ) + ax 2 ) < 0, x 0 and V (x) = x 0 g T (y)dy > 0 for x 0 [ ] α(x)x1 + β(x)x Choose g(x) = 2 where α, β, γ, δ to be determined γ(x)x 1 + δ(x)x 2 Farzaneh Abdollahi Nonlinear Control Lecture 4 21/70

22 To satisfy the symmetry req., we need β(x) + α x 2 x 1 + β x 2 x 2 = γ(x) + γ x 1 x 1 + δ x 1 x 2 V (x) = α(x)x 1 x 2 + β(x)x 2 2 aγ(x)x 1x 2 aδ(x)x 2 2 δ(x)x 2h(x 1 ) γ(x)x 1 h(x 1 ) To cancel the cross terms, let α(x)x 1 aγ(x)x 1 δ(x)h(x 1 ) = 0 V (x) = (aδ(x) β(x))x 2 2 γ(x)x 1h(x 1 ) For simplification, let δ(x) = δ = cte, γ(x) = γ = cte, β(x) = β = cte α(x) only depends on x 1 symmetry is satisfied if β = γ. [ aγx1 + δh(x g(x) = 1 ) + γx 2 γx 1 + δx 2 ] Farzaneh Abdollahi Nonlinear Control Lecture 4 22/70

23 By integration, we get V (x) = [ aγ γ where P = γ δ x1 0 = 1 2 aγx δ = 1 2 x T Px + δ ]. (aγy 1 + δh(y 1 ))dy 1 + x1 0 x1 0 x2 0 (γx 1 + δy 2 )dy 2 h(y)dy + γx 1 x δx 2 2 h(y)dy Choosing δ > 0, 0 < γ < aδ = V is p.d. & V is n.d. e.g., taking γ [ = akδ, 0 ] < k < 1 yields V (x) = δ ka 2 x T 2 ka x + δ x 1 ka 1 0 h(y)dy over D = {x R n b < x 1 < c} conditions of the theorem are satisfied x = 0 is asymptotically stable. Farzaneh Abdollahi Nonlinear Control Lecture 4 23/70

24 Region of Attraction For asymptotically stable Equ. pt.: How far from the origin can the trajectory be and still converges to the origin as t? Let φ(t, x) be the sol. of ẋ = f (x) starting at x 0. Then, the Region of Attraction (RoA) is defined as the set of all pts. x s.t. φ(t, x) = 0 lim t Lyap. fcn can be used to estimate the RoA: If there is a Lyap. fcn. satisfying asymptotic stability over domain D, and set Ω c = {x R n V (x) c} is bounded and contained in D all trajectories starting in Ω c remains there and converges to 0 at t. Under what condition the RoA be R n ( i.e., the Equ. pt. is globally asymptotically stable (g.a.s))? the conditions of stability theory must hold globally, i.e. D = R n. This not enough! for large c, the set Ωc should be kept bounded. i.e., reduction of V (x) should also result in reduction of x. Farzaneh Abdollahi Nonlinear Control Lecture 4 24/70

25 Region of Attraction Example: V (x) = x2 1 1+x x 2 2 It s clear that V (x) can get smaller, but x grows unboundedly Babashin-Krasovskii Theorem: Let x = 0 be an Equ. pt. of ẋ = f (x). Let V : R n R be a continuously differentiable fcn. s.t.: V (0) = 0 V (x) > 0, x 0 x = V (x) (i.e. it is radially unbounded) V < 0, x 0 then x = 0 is globally asymptotically stable Farzaneh Abdollahi Nonlinear Control Lecture 4 25/70

26 Example 6 : Globally Asymptotically Stable Reconsider Example 5 ( h(0) = 0, xh(x) > 0, x 0, x ( a, a)) but assume that xh(x) > 0 hold for all x 0. The Lyap. fcn: V (x) = δ [ 2 x T ka 2 ka ka 1 ] x1 x + δ h(y)dy 0 is p.d. x R 2 V (x) is radially unbounded. V = aδ(1 k)x2 2 akδx 1h(x 1 ) < 0, x R 2 origin is g.a.s. Important: Since the origin is g.a.s., then it must be the unique Equ. pt. of the system g.a.s. is not satisfied for multiple Equ. pt. problem such as pendulum. Farzaneh Abdollahi Nonlinear Control Lecture 4 26/70

27 Instability Theorem Chetaev s Theorem Let x = 0 be an Equ. pt. of ẋ = f (x). Let V : D R be a continuously differentiable fcn such that V (0) = 0 and V (x 0 ) > 0 for some x 0 with arbitrary small x 0. Define a set ν = {x B r V (x) > 0} where B r = {x R n x < r} and suppose that V (x) is p.d. in ν. Then, x = 0 is unstable. Example: ẋ 1 = x 1 + g 1 (x) ẋ 2 = x 2 + g 2 (x) where g i (x) k x 2 2 in a neighborhood D of origin The inequality implies, g i (0) = 0 = origin is an Equ. pt. Consider: V (x) = 1 2 (x 2 1 x 2 2 ) On the line x2 = 0, V (x) > 0. V (x) = x x2 2 + x 1g 1 (x) x 2 g 2 (x) 2 Since x1 g 1 (x) x 2 g 2 (x) x i g i (x) 2k x 3 2 V (x) x 2 2 2k x 3 2 = x 2 2 (1 2k x 2) Choosing r s.t. B r D and r < 1 2k = origin is unstable. i=1 Farzaneh Abdollahi Nonlinear Control Lecture 4 27/70

28 Invariance Principle Recall the pendulum example: V (x) = k which is n.s.d. ẋ 1 = x 2 ẋ 2 = g sinx 1 k l m x 2 m x 2 2 Lyap. theorem shows only stability. However, V is negative everywhere except at x2 = 0 where V = 0. To get V = 0, the trajectory must be confined to x 2 = 0 Now, from the model x2 (t) 0 = ẋ 1 0 = x 1 (t) = cte and x 2 (t) 0 = ẋ 2 0 = sinx 1 0 Hence, on the segment π < x 1 < π of x 2 = 0 line, the system can maintain V = 0 only at x = 0. Therefore, V (x) decrease to zero and x(t) 0 as t. The idea follows from LaSalle s Invariance Principle Farzaneh Abdollahi Nonlinear Control Lecture 4 28/70

29 Invariance Principle A set M is said to be a positively invariant set with respect to ẋ = f (x), if x(0) M = x(t) M, t 0 If a solution belongs to M at some time instant, then it belongs to M for all future time. Equ. points and limit cycle are invariant sets Also the set Ω = {x R n V (x) c} with V 0 x Ω is a positively invariant set. Farzaneh Abdollahi Nonlinear Control Lecture 4 29/70

30 Lasalle s Theorem: Let Ω be a compact set with property that every solution of ẋ = f (x) starting in Ω remains in Ω for all future time. Let V : Ω R be a continuously differentiable fcn s.t. V (x) 0 in Ω. Let E be the set of all pts in Ω where V (x) = 0 Let M be the largest invariant set in E. Then, every sol. starting in Ω approaches M as t Unlike Lyap. theorem, Lasalle s theorem does not require V (x) to be p.d Only Ω should be bounded If V is p.d. Ω = {x R n V (x) c} is bounded for suff. small c If V is radially unbounded Ω is bounded for all c no matter V is p.d. or not To show a.s. of the origin show largest invariant set in E is the origin. Show that no solution can stay forever in E other than x = 0. Farzaneh Abdollahi Nonlinear Control Lecture 4 30/70

31 Lasalle s Theorem: Farzaneh Abdollahi Nonlinear Control Lecture 4 31/70

32 Barbashin and Krasovskii Corollaries Corollary 1: Let x = 0 be an Equ. pt of ẋ = f (x). Let V : D R be a continuously differentiable p.d. fcn on a domain D containing the origin x = 0, s.t. V (x) 0 in D. Let S = {x D V = 0} and suppose that no solution can stay identically in S, other than the trivial solution x(t) = 0. Then, the origin is a.s. Corollary 2: Let x = 0 be an Equ. pt. of ẋ = f (x). Let V : R n R be a continuously differentiable, radially unbounded, p.d. fcn s.t. V (x) 0 x R n. Let S = {x R n V = 0} and suppose that no solution can stay in S forever except x = 0. Then, the origin is g.a.s. Farzaneh Abdollahi Nonlinear Control Lecture 4 32/70

33 Example 6: Consider ẋ 1 = x 2 ẋ 2 = g(x 1 ) h(x 2 ) where g(.) & h(.) are locally Lip. and satisfy g(0) = 0, yg(y) > 0 y 0, y ( a, a) h(0) = 0, yh(y) > 0 y 0, y ( a, a) The system has an isolated Equ. pt. at origin. Let V (x) = 1 x1 2 x g(y)dy 0 D = {x R 2 a < x i < a} = V (x) > 0 in D V = g(x 1 )x 2 + x 2 ( g(x 1 ) h(x 2 )) = x 2 h(x 2 ) 0 Thus, V is n.s.d. and the origin is stable by Lyap. theorem Farzaneh Abdollahi Nonlinear Control Lecture 4 33/70

34 Example 6: Using Lasalle s theorem, define S = {x D V = 0} V = 0 = x 2 h(x 2 ) = 0 = x 2 = 0, since a < x 2 < a Hence S = {x D x2 = 0}. Suppose x(t) is a traj. S t x 2 (t) 0 = ẋ 1 0 = x 1 (t) = c, where c ( a, a). Also x 2 (t) 0 = ẋ 2 0 = g(c) = 0 = c = 0 Only solution that can stay in S t 0 is the origin = x = 0 is a.s. Now, Let a = and assume g satisfy: y 0 g(z)dz as y. The Lyap. fcn V (x) = 1 2 x x 1 0 g(y)dy is radially unbounded. V 0 in R 2 and note that S = {x R 2 V = 0} = {x R 2 x 2 = 0} contains no solution other than origin = x = 0 is g.a.s. Farzaneh Abdollahi Nonlinear Control Lecture 4 34/70

35 Summary Lyapunov Direct Method:The origin of an autonomous system ẋ = f (x) is stable if there is a continuously differentiable, p.d. function V (x) s.t. V (x) is n.s.d., and] is asymptotically if V (x) is n.d. V (x) is p.d./p.s.d. iff λ i {P} > 0/λ i {P} 0 iff all leading principle minors of P are positive / non-negative. Variable Gradient Method: To find a Lyap fcn: Choose g(x) s.t. 1. g i x j = g j x i 2. g T (x)f (x) is n.d. (n.s.d) 3. V (x) = x1 g 0 1(y 1, 0,..., 0)dy 1 + x 2 is P.d. g 0 2(x 1, y 2, 0,..., 0)dy x n g 0 n(x 1, x 2,..., y n)dy n Babashin-Krasovskii Theorem: Let x = 0 be an Equ. pt. of ẋ = f (x). Let V : R n R be a continuously differentiable fcn. s.t.: V (0) = 0, V (x) > 0, x 0, x = V (x) (i.e. it is radially unbounded), V < 0, x 0 then x = 0 is globally asymptotically stable Farzaneh Abdollahi Nonlinear Control Lecture 4 35/70

36 Summary Another method for study a.s is defined based on Lasalle Theorem: Corollary 1: Let x = 0 be an Equ. pt of ẋ = f (x). Let V : D R be a continuously differentiable p.d. fcn on a domain D containing the origin x = 0, s.t. V (x) 0 in D.Let S = {x D V = 0} and suppose that no solution can stay identically in S, other than the trivial solution x(t) = 0. Then, the origin is a.s. The origin is g.a.s. if D = R n, and V (x) is radially unbounded. Farzaneh Abdollahi Nonlinear Control Lecture 4 36/70

37 Invariance Principle Lasalle s theorem can also extend the Lyap. theorem in three different directions 1. It gives an estimate of the RoA not necessarily in the form of Ω c = {x R n V (x) c}. The set can be any positively invariant set which leads to less conservative estimate. 2. Can determine stability of Equ. set, rather than isolated Equ. pts. 3. The function V (x) does not have to be positive definite. Example 7: shows how to use Lasalle s theorem for system with Equ. sets rather than isolated Equ. pts A simple adaptive control problem: with the adaptive control law ẋ = ax + u a unknown u = kx; k = γx 2, γ > 0 Farzaneh Abdollahi Nonlinear Control Lecture 4 37/70

38 Example 7: Let x 1 = x, x 2 = k, we get: ẋ 1 = (x 2 a)x 1 ẋ 2 = γx 2 1 The line x 1 = 0 is an Equ. set Show that the traj. of closed-loop system approaches this set as t i.e. the adaptive system regulates y to zero (x 1 0 as t ). Consider the Lyap. fcn candidate: V (x) = 1 2 x γ (x 2 b) 2 where b > a. V = x 2 1 (b a) 0 The set Ω = {x R n V (x) c} is a compact positively invariant set Farzaneh Abdollahi Nonlinear Control Lecture 4 38/70

39 Example 7: V (x) is radially unbounded = Lasalle s theorem conditions are satisfied with the set E as E = {x Ω x 1 = 0} Since any pt on x 1 = 0 line is an Equ. pt, E is an invariant set: M = E. Hence, every trajectory starting in Ω approaches E as t, i.e. x 1 (t) 0 as t. V is radially unbounded = the result is global Note that in the above example the Lyapunov function depends on a constant b which is required to satisfy b > a But it is not known we may not know the constant b explicitly, but we know that it always exists. This highlights another feature of Lyapunov s method: In some situations, we may be able to assert the existence of a Lyapunov function that satisfies the conditions, even though we may not explicitly know that function. Farzaneh Abdollahi Nonlinear Control Lecture 4 39/70

40 Linear System and Linearization Given ẋ = Ax, the Equ. pt. is at origin It is isolated iff det A 0, System has an Equ. subspace if det A = 0, the subspace is the null space of A. The linear system cannot have multiple isolated Equ. pt. since Linearity requires that if x1 and x 2 are Equ. pts., then all pts. on the line connecting them should also be Equ. pts. Theorem: The Equ. pt. x = 0 of ẋ = Ax is stable iff all eigenvalues of A satisfy Re{λ i } 0 and every eigenvalue with Re{λ i } = 0 and algebric multiplicity q i 2, rank(a λ i I ) = n q i, where n is dimension of x. The Equ. pt. x = 0 is globally asymptotically stable iff Reλ i < 0. Farzaneh Abdollahi Nonlinear Control Lecture 4 40/70

41 Linear Systems and Linearization When all eigenvalues of A satisfy Reλ i < 0, A is called a Hurwitz matrix. Asymptotic stability can be verified by using Lyapunov s method : Consider a quadratic Lyap. fcn candidate: where V (x) = x T P x, P = P T > 0 V = ẋ T Px + x T Pẋ, x T (A T P + PA)x x T Q x A T P + P A = Q, Q = Q T Lyapunov Equation If Q is p.d., then we conclude that x = 0 is g.a.s. We can proceed alternatively as follows: Start by choosing Q = Q T, Q > 0, then solve the Lyap. eqn. for P. If P > 0, then x = 0 is g.a.s. Farzaneh Abdollahi Nonlinear Control Lecture 4 41/70

42 Linear Systems and Linearization Theorem: A matrix A is a stable matrix, i.e. Re λ i < 0 iff for every given Q = Q T > 0, P = P T > 0 that satisfies the Lyap. eq. Moreover, if A is a stable matrix, then P is unique. Example 8: [ ] 0 1 A = 1 1 [ ] 1 0 Let Q = = Q 0 1 T > 0 [ ] P11 P denote P = 12 = P P 12 P T > 0 22 The Lyap. eq. A T P + P A = Q becomes 2 P 12 = 1 P 11 P 12 + P 22 = 0 2 P 12 2P 22 = 1 Farzaneh Abdollahi Nonlinear Control Lecture 4 42/70

43 Linear Systems and Linearization P 11 P 12 P 22 = = P 11 P 12 P 22 = (1) Let P = P T = [ ] > 0 = x = 0 is g.a.s Remark: Computationally, there is no advantages in computing the eigenvalues of A over solving Lyap. eqn. Farzaneh Abdollahi Nonlinear Control Lecture 4 43/70

44 Linear Systems and Linearization Consider ẋ = f (x) where f : D R n, D R n, is continuously diff. Let x = 0 is in the interior of D and f (0) = 0. In a small neighborhood of x = 0, the nonlinear system ẋ = f (x) can be linearized by ẋ = Ax. Proof: Ref. Khalil s book Theorem (Lyapunov s First Method): Let x = 0 be an Equ. pt. for ẋ = f (x) where f : D R n is continuously differentiable and D is a nghd of origin. Let A = f x x=0, then 1. x = 0 is a.s. if Reλ i < 0, i = 1,..., n 2. x = 0 is unstable if Reλ i > 0, for one or more eigenvalues Farzaneh Abdollahi Nonlinear Control Lecture 4 44/70

45 Linear Systems and Linearization Example 9: ẋ = ax 3 Linearization about x = 0 yields: A = f x = 3ax 2 x=0 = 0 x=0 Linearization fails to determine stability If a < 0, x = 0 is a.s. To see this, let V (x) = x 4 = V = 4x 3 ẋ = 4ax 6 If a > 0, x = 0 is unstable If a 0, x = 0 is stable, starting at any x, remains in x Example 10: ẋ 1 = x 2 ( g ẋ 2 = l ) ( ) k sin x 1 m x 2 Farzaneh Abdollahi Nonlinear Control Lecture 4 45/70

46 Linear Systems and Linearization Linearization about 2 Equ. pts. (0, 0) & (π, 0): A = f x = [ f1 x 1 f 2 x 2 f 2 f 1 x 1 x 2 ] [ 0 1 = g l cos x 1 k m ] At (0, 0) A = [ 0 1 g l k m ], λ 1,2 = 1 2 k m ± 1 2 ( k m )2 4 g l g, k, l, m > 0 = Re(λ1, λ 2 ) < 0 = x = 0 is a.s. If k = 0 = Re(λ1, λ 2 ) = 0 = eigenvalues on jω axis,rank(a λ i I ) 0 stability cannot be determined. At (π, 0), change the variable to z 1 = x 1 π, z 2 = x [ ] A = g, λ 1,2 = 1 k 2 m ± 1 2 ( k m )2 + 4 g l l k m g, k, l, m > 0 = there is one eigenvalue in the open right-half plane = x = 0 is unstable. Farzaneh Abdollahi Nonlinear Control Lecture 4 46/70

47 Example: Robot Manipulator Dynamics: M(q) q + C(q, q) q + B q + g(q) = u (2) where M(q) is the n n inertia matrix of the manipulator C(q, q) q is the vector of Coriolis and centrifugal forces g(q) is the term due to the Gravity B q is the viscous damping term u is the input torque, usually provided by a DC motor Objective: To regulate the joint position q around desired position q d. A common control strategy PD+Gravity: u = K P q K D q + g(q) where q = q d q is the error between the desired and actual position K P and K D are diagonal positive proportional and derivative gains. Farzaneh Abdollahi Nonlinear Control Lecture 4 47/70

48 Example: Robot Manipulator Consider the following Lyap. fcn candidate: V = 1 2 qt M(q) q qt K P q The first term is the kinetic energy of the robot and the second term accounts for artificial potential energy associated with virtual spring in PD control law (proportional feedback K p q) Physical properties of a robot manipulator: 1. The inertia matrix M(q) is positive definite 2. The matrix Ṁ(q) 2C(q, q) is skew symmetric V is positive in R n except at the goal position q = q d, q = 0 V = q T M(q) q qt Ṁ(q) q q T K P q Substituting M(q) q from (2) into the above equation yields V = q T (u C(q, q) q B q g(q)) qt Ṁ(q) q q T K P q = q T (u B q K P q g(q)) qt (Ṁ(q) 2C(q, q)) q Farzaneh Abdollahi Nonlinear Control Lecture 4 48/70

49 Example: Robot Manipulator V = q T (u B q K P q g(q)) where Ṁ 2C is skew symmetric q T (Ṁ(q) q C(q, q)) q = 0 Substitute PD control law for u, we get: V = q T (K D + B) q 0 (3) The goal position is stable since V is non-increasing Use the invariant set theorem: Suppose V 0, then (3) implies that q 0 and hence q 0 From Equ. of motion (2) with PD control, we have M(q) q + C(q, q) q + B q = K P q K D q we must then have 0 = K P q which implies that q = 0 V is radially unbounded. Global asymptotic stability is ensured. Farzaneh Abdollahi Nonlinear Control Lecture 4 49/70

50 Example: Robot Manipulator In case, the gravitational terms is not canceled, V is modified to: V = q T ((K D + B) q + g(q)) The presence of gravitational term means PD control alone cannot guarantee asymptotic tracking. Assuming that the closed loop system is stable, the robot configuration q will satisfy K P (q d q) = g(q) The physical interpretation of the above equation is that: The configuration q must be such that the motor generates a steady state holding torque K P (q d q) sufficient to balance the gravitational torque g(q). the steady state error can be reduced by increasing K P. Farzaneh Abdollahi Nonlinear Control Lecture 4 50/70

51 Control Design Based on Lyapunov s Direct Method Basically there are two approaches to design control using Lyapunov s direct method Choose a control law, then find a Lyap. fcn to justify the choice Candidate a Lyap. fcn, then find a control law to satisfy the Lyap. stability conditions. Both methods have a trial and error flavor In robot manipulator example the first approach was applied: First a PD controller was chosen based on physical intuition Then a Lyap. fcn. is found to show g.a.s. Farzaneh Abdollahi Nonlinear Control Lecture 4 51/70

52 Control Design Based on Lyapunov s Direct Method Example: Regulator Design Consider the problem of stabilizing the system: ẍ ẋ 3 + x 2 = u In other word, make the origin an asymptotically stable Equ. pt. Recall the example: ẋ 1 = x 2 ẋ 2 = g(x 1 ) h(x 2 ) where g(.) & h(.) are locally Lip. and satisfy g(0) = 0, yg(y) > 0 y 0, y ( a, a) h(0) = 0, yh(y) > 0 y 0, y ( a, a) Asymptotic stability of such system could be shown by selecting the following Lyap. fcn: V (x) = 1 x1 2 x g(y)dy 0 Farzaneh Abdollahi Nonlinear Control Lecture 4 52/70

53 Example: Regulator Design Let x 1 = x, x 2 = ẋ. The above example motivates us to select the control law u as u = u 1 (ẋ) + u 2 (x) where ẋ(ẋ 3 + u 1 (ẋ)) < 0 for ẋ 0 x(u 2 (x) x 2 ) < 0 for x 0 The globally stabilizing controller can be designed even in the presence of some uncertainties on the dynamics: ẍ + α 1 ẋ 3 + α 2 x 2 = u where α 1 and α 2 are unknown, but s.t. α 1 > 2 and α 2 < 5 This system can be globally stabilized using the control law: u = 2ẋ 3 5(x + x 3 ) Farzaneh Abdollahi Nonlinear Control Lecture 4 53/70

54 Estimating Region of Attraction Sometimes just knowing a system is a.s. is not enough. At least an estimation of RoA is required. Example: Occurring fault and finding critical clearance time Let x = 0 be an Equ. pt. of ẋ = f (x). Let φ(t, x) be the sol starting at x at time t=0. The Region Of Attraction (RoA) of the origin denoted by R A is defined by: R A = {x R n φ(t, x) 0 as t } Lemma: If x = 0 is an a.s. Equ. pt. of ẋ = f (x), then its RoA R A is an open, connected, invariant set. Moreover, the boundary of RoA, R A, is formed by trajectories of ẋ = f (x). one way to determine RoA is to characterize those trajectories that lie on R A. Farzaneh Abdollahi Nonlinear Control Lecture 4 54/70

55 Example: Van-der-Pol Dynamics of oscillator in reverse time ẋ 1 = x 2 ẋ 2 = x 1 + (x 2 1 1)x 2 The system has an Equ. pt at origin and an unstable limit cycle. The origin is a stable focus it is a.s. Farzaneh Abdollahi Nonlinear Control Lecture 4 55/70

56 Example: Van-der-Pol Checking by linearizaton method A = f [ 0 1 x x=0 1 1 ] λ = 1/2 ± j 3/2 Re λ i < 0 Clearly, RoA is bounded since trajectories outside the limit cycle drift away from it R A is the limit cycle Farzaneh Abdollahi Nonlinear Control Lecture 4 56/70

57 Example 11: ẋ 1 = x 2 ẋ 2 = x x 1 3 x 2 There are 3 isolated Equ. pts. (0, 0), ( 3, 0), ( 3, 0). (0, 0) is a stable focus, the other two are saddle pts. Origin is a.s. and other two are unstable (follows from linearization). stable trajectories of the saddle points form two separatrices that are R A RoA is unbounded Farzaneh Abdollahi Nonlinear Control Lecture 4 57/70

58 Example 11: Recall the example: ẋ 1 = x 2 ẋ 2 = h(x 1 ) ax 2 V = δ [ 2 x T ka 2 ka ka 1 ] x1 x + δ hdy 0 Let V = 1 2 x T [ = 3 4 x x x 1x x 2 2 We get: V = 1 2 x 1 2 ( x 1 2 ] x + x 1 ( 0 y 1 3 y 3) dy ) 1 2 x 2 2 Define D = {x R 2 3 < x 1 < 3} V (x) > 0 and V (x) < 0 in D {0}, From the phase portrait = D is not a subset of R A. Tell me why?!! Farzaneh Abdollahi Nonlinear Control Lecture 4 58/70

59 Estimating RoA Traj starting in D move from one Lyap. surface to V (x) = c 1 to an inner surface V (x) = c 2 with c 2 < c 1. However, there is no guarantee that the traj. will remain in D forever. Once, the traj leaves D, no guarantee that V remains negative. This problem does not occur in R A since R A is an invariant set. The simplest estimate is given by the set Ω c = {x R n V (x) c} where Ω c isbounded and connected and Ω c D Note that {V (x) c} may have more than one component, only the bounded component which belong to D is acceptable. Example: If V (x) = x 2 /(1 + x 4 ).and D = { x < 1} The set {V (x) 1/4} has two components { x 2 3} and { x 2 + 3} only { x 2 3} is acceptable. Farzaneh Abdollahi Nonlinear Control Lecture 4 59/70

60 Estimating RoA To find RoA, first we need to find a domain D in which V is n.d. Then, a bounded set Ω c D shall be sought We are interested in largest set Ω c, i.e. the largest value of c since Ω c is an estimate of R A. V is p.d. everywhere in R 2. If V (x) = x T Px, let D = {x R 2 x r}. Once, D is obtained, then select Ω c D by c < min V (x) x =r In words, the smallest V (x) = c which fits into D. Since We can choose To enlarge the estimate of R A x T Px λ min (P) x 2 c < λ min (P)r 2 = find largest ball on which V is n.d. Farzaneh Abdollahi Nonlinear Control Lecture 4 60/70

61 Example 12: ẋ 1 = x 2 ẋ 2 = x 1 + (x1 2 1)x 2 [ ] From the linearization f 0 1 x x=0 = is stable 1 1 Taking Q = I and solve the Lyap. [ equation: ] PA + A T P = I = P =.5 1 λ min (P) = 0.69 V = (x x2 2) (x 1 3x 2 2x1 2x 2 2) x x 1 x 1 x 2 x 1 2x 2 x x 4 2 where x 1 x 2, x 1 x 2 x 2 2 /2, x1 2x 2 5 x 2 V is n.d. on a ball D of radius r 2 = 2/ 5 = c < = Farzaneh Abdollahi Nonlinear Control Lecture 4 61/70

62 Example 12: To find less conservative estimate of Ω c : Let x 1 = ρ cosθ, x 2 = ρ sinθ V = ρ 2 + ρ 4 cos 2 θsinθ (2sinθ cosθ) ρ 2 + ρ 4 cos 2 θsinθ 2sinθ cosθ ρ 2 + ρ 4 (.3849)(2.2361) ρ ρ 4 < 0 for ρ 2 < c =.8 < =.801 Thus the set: Ω c = {x R 2 V (x).8} is an estimate of R A. Farzaneh Abdollahi Nonlinear Control Lecture 4 62/70

63 Example 12: A lesser conservative estimation of RoA: plot the contour of V = 0 plot V (x) = c for increasing c to find largest c where V < 0 The c obtained by this method is c = Farzaneh Abdollahi Nonlinear Control Lecture 4 63/70

64 Example 13: ẋ 1 = 2x 1 + x 1 x 2 ẋ 2 = x 2 + x 1 x 2 There are two Equ. pts., (0, 0), (1, 2). [ ] 0 1 At (1, 2) A = = unstable ( λ 2 0 1,2 = ± 2 ) (saddle pt.) [ ] 2 0 At (0, 0) A = = a.s. 0 1 [ 1 ] Taking Q = I and solving Lyap Eq. A T P + PA = I P = The Lyap. fcn is V (x) = x T Px We have V = (x x 2 2 ) (x 2 1 x 2 + 2x 1 x 2 2 ) Find largest D s.t. V is n.d. in D. Let x 1 = ρ cosθ, x 2 = ρ sinθ Farzaneh Abdollahi Nonlinear Control Lecture 4 64/70

65 Example 13: V = ρ 2 + ρ 3 cosθsinθ (sinθ + 12 ) cosθ ρ ρ3 sin2θ sinθ cosθ 5 ρ ρ3 < 0 for ρ < 4 5 Since λ min (P) = 1 4 Thus the set: =, we choose c =.79 < 1 4 ( 4 5 ) 2 =.8 Ω c = {x R 2 V (x).79} R A. Estimating RoA by the set Ω c is simple but conservative Alternatively Lasalle s theorem can be used. It provides an estimate of R A Farzaneh Abdollahi Nonlinear Control Lecture 4 65/70

66 Example 14: ẋ 1 = x 2 ẋ 2 = 4(x 1 + x 2 ) h(x 1 + x 2 ) where h : R R s.t. h(0) = 0, &xh(x) 0 x 1 Consider the Lyap fcn candidate: V (x) = x T [ ] x = 2x x 1 x 2 + x 2 2 Then V = 2x1 2 6(x 1 + x 2 ) 2 2(x [ 1 + x 2 )h(x ] 1 + x 2 ) 8 6 2x1 2 6(x 1 + x 2 ) 2 = x T x, x x 2 1 V is n.d. in the set G = {x R 2 x 1 + x 2 1}. (0, 0) is a.s., to estimate R A, first do it from Ω c. Farzaneh Abdollahi Nonlinear Control Lecture 4 66/70

67 Example 14: Find the largest c s.t. Ω c G. Now, c is given by c = min V (x) or x 1 +x 2 =1 { } c = min min V (x), min V (x) x 1 +x 2 =1 x 1 +x 2 = 1 The first minimization yields { min V (x) = min 2x 2 x 1 +x 2 =1 x 1 + 2x 1 (1 x 1 ) + (1 x 1 ) 2} = 1 and 1 min V (x) = 1 x 1 +x 2 = 1 Hence, Ω c with c = 1 is an estimate of R A. A better (less conservative) estimate of R A is possible. Farzaneh Abdollahi Nonlinear Control Lecture 4 67/70

68 Example 14: The key point is to observe that traj inside G cannot leave it through certain segment of the boundary x 1 + x 2 = 1. Let σ = x 1 + x 2 = G is given by σ = 1 and σ = 1 We have d dt σ2 = 2σ( x 1 + x 2 ) = 2σx 2 8σ 2 2σh(σ) 2σx 2 8σ 2, σ 1 On the boundary σ = 1 = dσ2 dt 2x x 2 4 Hence, the traj on σ = 1 for which x 2 4 cannot move outside the set G since σ 2 is non-increasing Similarly, on the boundary σ = 1 we have dσ 2 dt 2x x 2 4 Hence, the traj on σ = 1 for which x 2 4 cannot move outside the set G. Farzaneh Abdollahi Nonlinear Control Lecture 4 68/70

69 Example 14: To define the boundary of G, we need to find two other segments to close the set. We can take them as the segments of Lyap. fcn surface Let c 1 be s.t. V (x) = c 1 intersects the boundary of x 1 + x 2 = 1 at x 2 = 4 and let c 2 be s.t. V (x) = c 2 intersects the boundary of x 1 + x 2 = 1 at x 2 = 4 Then, we define V (x) = min c 1, c 2, we have c 1 = V (x) x1 = 3 x 2 = 4 = 10 &c 2 = V (x) x 1 = 3 x 2 = 4 = 10 Farzaneh Abdollahi Nonlinear Control Lecture 4 69/70

70 Example 14: The set Ω is defined by Ω = {x R 2 V (x) 10 & x 1 + x 2 1} This set is closed and bounded and positively invariant. Also, V is n.d. in Ω since Ω G = Ω R A. Farzaneh Abdollahi Nonlinear Control Lecture 4 70/70