Proceeding of the 2004 American Control Conference FrM04.

4
Boston. Massachusetts J u n e 30 -July 2,2004

A Globally Stabilising Controller under Saturated Input for

Linear Planar Systems with One Unstable Pole
J-Y.Favez, Ph. Mullhaupt, B. Shivasan, D.' Bonvin '

Laboratoire d'Automatique, Ecole Polytechnique FBdBrale d e Lausanne,

CH-1015Lausanne, Switzerland
email: jean-yves.favez@epfl.ch; fax: +41 21 693 25 74

AbsfrucGControllers for planar systems with one stable Local performance - Enforcing the desired perfor-
and one unstable pole under saturated input are considered. mance locally around the origin: For systems with one
The requirements are: (i) global stability, i.e. the region of stable and one unstable pole, a simple way of obtain-
attraction is the null controllable region, (ii) possibility of
enforcing any desired performance around the origin, and (iii) ing global stability is to feed back only the unstable
no chattering of the control signal in the presence of noise. A state. However, such a controller will have poor local
simple continuous noulinear state feedback controller is pro- performance due to absence of feedback on the stable
posed that satisfies all these requirements. The performance mode. The proposed controller should guarantee, at
of the proposed controller is compared in simulation with that least locally, the desired temporal performance.
of classical controllers such as linear state feedback and time
optimal controllers. Absence of chattering in the control signal in the
Keywords: Region of Attraction, Saturated Input, Con- presence of noise: From the point of view of the
tinuous Nonlinear Control, Unstable Planar Systems. above two requirements, the optimal or near optimal
switching strategies are the methods of choice. The
I. INTRODUCTION controllers have the maximum region of attraction and
Linear systems with bounded inputs have been widely provide excellent time performance, not only locally,
studied in the literature [9], [6],[4]. This type of studies is but even globally. However, the main problem is that
important since, in most practical situations, the range of the control signal is chattering in the presence of
inputs is in fact limited. noise. Absence of chattering is a requirement for the
In this paper, the control of single-input linear planar sought controller.
systems (systems with 2 states) with saturated linear and
In what follows various standard controllers proposed in
nonlinear feedback will be considered. Especially, the em-
the literature are analysed from the perspective of control
phasis will be on systems with one stable and one unstable
of linear planar systems with one stable and one unstable
pole. The three main requirements addressed here are as
pole.
follows:
1) Global stability: Two important concepts pertaining i) Standard linear state feedback controller where both
to these systems have to be distinguished. First is the states are fed hack [3]: Requirement 1 is not satisfied.
null controllable region C, i.e. the region in state space ii) Linear state feedback controllers where only the un-
where there exists an open-loop input that can steer stable state is fed back [l], [PI: Requirement 2 is not
the system to the origin [I], [4], [5], [8]. The second satisfied since the stable state is left to follow its own
is the region of attraction A with a given controller , dynamics.
i.e. the region in state space from which the closed- iii) Optimal or near-time optimal VSC controller [7],[XI:
loop system asymptotically reaches the origin [I], [4]. Requirement 3 is not verified in presence of noise and
A controller is globally stabilising when A = C. disturbances.
For semi-stable planar systems (both poles have non- iv) Controller switching from controller (ii) to controller
positive real parts), C = W2 and A = C = W2 for any (i) [4], [5]: When the state reaches an invariant non-
linear state feedback which makes the origin globally saturated region, the controller switches from type (ii)
asymptotically stable in the absence of saturation to (i). The problem is that the invariant region could
[9]. However, for systems with one stable and one be very small.
unstable pole, C C W2 and A = C can be either The properties of these controllers are summarised in Ta-
achieved with an optimal or near optimal variable ble I. The controllers available in the literature cannot
structure control (VSC) [7],[SI or using a linear state satisfy one or more of the requirements stated above. Thus,
feedback controller where only the unstable state is a controller is proposed that meets all the aforementioned
fed back [I], [8]. For anti-stable systems (both poles requirements. The idea is to have a controller of type (iv)
have positive real parts), C c W2 and the only way with a continuous switching from controller type (ii) to
to obtain A = C is through optimal or near optimal controller type (i). This way, the problem of chattering
switching [4], [5]. is avoided and the region of attraction is the whole null
0-7803-8335-4/041$17.00 0 2 0 0 4 AACC 4681
 Definition 1 Let @(t,50)denote the state of (3) at time t,
starting with the initial condition xo at t = 0. 7he region of
Global Local NO attraction of the stable equilibrium point 2.0 = 0 is defined
stability performance chattering
as:
= 0} .
A = {z : Jis@(t,z) (4)
Feedback unstable stale I[ J I - I J-
mal controller II J I J The boundary of A is denoted by aA.

TABLE I C. Manifoldi
PROPERTIES OF THE STATE OF THE ART CONTROLLERS. Define the following hyperplanes and manifolds (refer to
The paper is organised as follows. In Section 11, delini-
tions and terms used in this paper are introduced. Section 111
..
Figure 1 for illustration):
aco = {z : ii = 0)
provides the main idea behind the structure of the proposed
continuous nonlinear controller. In Section IV,the global
..ac+ = {z : ii = I), ac- = tz : ii = -1)
ac+ = tz : z1 = I), ac- = {Z ; z1= -1)
stabilisation in the null controllable region is discussed.
Section V compares the proposed controller against the
. as+ = {z : limt,, @(t,z)= z,+}
8s- = {z : limt,, @(t,z)= l e - )
The hyperplanes aC+ and aC- are the boundaries of the
standard approaches. Conclusions are drawn in Section VI.
region C where the control is not saturated and aC0 is the
11. PRELIMINARIES hyperplane of zero control. The hyperplanes aC+ and aC-
A. Linear planar system with input saturation are the boundaries of null controllable region C [l], [PI.
Consider a single-input second-order linear system with aS denote the stable manifolds of the saddle points. All
manifolds have two branches, one on either side of the
a stable and an unstable pole. Upon state transformation,
the system can be written as: saddle points.
111. CONTINUOUS NONLINEAR STATE FEEDBACK
CONTROLLER
Consider the controller
where, z E R2 is the state vector, U the input, A and
b appropriate matrices, and X1, A2 the eigenvalues of the ii(z) = f1z1 +k ( z ) f z z z U = sat(ii), (5)
open-loop system. Assume that A1 > 0 and XZ < 0. The where f = [ fi fz ] E 1' and k(z): RZ -+ R. Assume
symmetric saturation function with unity saturation level that f has been chosen to get the desired performance
will be used
(Requirement 2) of the closed-loop system near the origin.
-1 if s < - 1 If we set k(z) = 1, then ( 5 ) is a linear state feedback
sat(s) = s
1 { if - 1 5 s 5 1
if s > l
With saturated state feedback, the closed-loop system is
(2) controller. If k ( z ) = 0, then the stable state is not fed
back leading to A = C [l]. Here, a continuous switching is
introduced by choosing:
k ( z ) = 1 - b11, (6)
j. = A s + b sat(ii(z)), (3)
where 0 < k(z) 5 1 since within the null controllable
where ii(z): W2 -+ W is the linear or nonlinear control law region 1511 < 1.
in function of the states. The idea behind this nonlinear controller is as follows. If
11 I I0, then k ( z ) % 1 which implies that the controller is
B. Equilibrium points and region of atfraction
approximately the linear state feedback ii zz flzl+fizz. In
Assume that the feedback ii(z) is a global stabiliser this case, the controller concentrates on local performance
for (1). Then, system (3) has three equilibrium points. (Requirement 2). On the contrary, if the unstable state
This conclusion, which regards an open-loop system with approaches the boundary of the null controllable region C,
one stable and one unstable pole, does not apply to all z1 U f l and k(z) z 0. This implies that the controller is
other open-loop pole configurations (both poles stable or approximately the linear state feedback ii % flzl,where it
unstable) where the origin is the unique equilibrium point focuses on the stabilisation of the unstable state and global
[I]. stability (Requirement 1). Since the controller (5)-(6) is a
continuous one, chattering is avoided and Requirement 3 is
Theorem 1 [I] Let the feedback ii be a global stabiliserfor also fulfilled.
( I ) at the origin. Then, the closed-loop system (3) has three Since f stabilises the system locally, it satisfies the
equilibrium points: x,+ = A-'b = [ 1 11'. 2,- = Hunvitz stability conditions:
T
-A-'b = [ -1 -1 ] , and xeo = 0. Of these, xe0 is 1) & ( I + f l ) + W l + fd < 0
stable, while the orher two are saddle points. 2) X1X2(1+ fl + fd > 0, (1 + fl + fd < 0
4682
Also the condition that the system is stable at k ( z ) = 0 Proof: Consider the input ( 5 ) along the manifold aC+:
+
implies (1 fl) < 0. U = sat(f1) = -1, since f1 < -1. Then, (8) along aC+
To picture the regions where the control is saturated and reads:
where it is not, consider the function g,j : ( - 1 , l ) -+ R that
describes z2 as a function of 2 1 for a given 4 E [ - l , l ] : $1 - 1) = 0
= X1(21 (9)
$2 = XZ(Z2 - 1) (10)
(7) Thus, z1 stays at 1, and since Xz < 0, z2 converges to 1 as
well. Thus, aC+ forms the set of all points that converge
With this function, it is possible to express aL0, aL+ and to the equilibrium point z.+. So, as+ = aC+. A similar
8L- by setting 4 = 0, 4 = 1 and 4= -1, respectively: proof can be written for as- = aC-.

Lemma 1 Let X = ( - 1 , l ) x [-l, l] and D = [-1 + e , 1 -

1 x [-1,1], wifh c > 0 chosen such that I U ( Z ) ~ = lfor all
2 E ( X \ D ) . Then. all frajectories starting in C will enter
the compact invariant set D.
aL- =
{ 2 E

Figure 1 illustrates XO,

(-1,l) x w: 22 =
-1 - fl.1
fdl- 11.1)
8L+,aL-,the boundary of the
I Proof: Consider the subset X+ = ( - 1 , l ) x (1, w].
Within this set,

null controllable region &?+and aC-, and the equilibrium iz = Xz(2z +U)<0
saddle nodes ze+ and ze-. since XZ < 0, 2 2 > 1, and 1 ~ 5 1 1 due to saturation.
So, all trajectories starting in X, will leave X+.A similar
argument can be provided for X+ = (-1,l) x [-w, -1).
So, since C is invariant, all trajectories starting in C will
reach X = C \ ( X + U X - ) = ( - 1 , l ) x [-1,1].
Consider the subset Y- = (-1, -1 + e ) x [-1,1], with
E > 0 chosen such that U
(
.) = 1 for all 2 E Y-.So, within
this set
$1 = X l ( 2 1 + U ) >0
since XI > 0, 51 > -1, and U = 1. Thus, eliminating
Y- and its counterpart Y+ = ( 1 , l - e ) x [-I, 11 gives
the compact invariant set D = X \ (Y+ U Y - ) , So, every
trajectory starting in X reaches D.
The proof is complete since every trajectory starting in
C reaches X and every trajectory starting in X reaches D.
Thus, D is invariant.

Fig. 1. Illustration of BCo, B&+, 8 L . the boundary of the null
controllable region Bc+ and aC-. the stable manifolds as+ and as-,
and the equilibrium saddle " d e s z.+ and z e - ,
IV. ANALYSIS
OF THE PROPOSED
CONTROLLER
Consider the closed-loop system with (3), ( 5 ) and (6) :
i=Azfbsat(fizl+fizz(l-13rll))
Lemma 2 I f X l + XZ > 0 or (fz - ji) 5 0, fhen the non-
sahrrated region L is an invarianf set for (8).
Pmof: The lemma is proved by showing that the
vector field of (8) points into the non-saturated region L
for every point along the manifolds aL+ and aL-. Since
aL+ and aL- are symmetric it suffices to prove it only for
one manifold (X+,4 = 1).
(8) The slope of the tangent of aL+,SI,is given by:

It was shown in [3] that the stable manifolds are the

boundaries of the region of attraction A. In what follows, it
will be shown that, with controller (5) the stable manifolds Note that SI < 0 for fz < 0 and s1 > 0 for fi > 0.
are in fact the boundaries of the null controllable region C. Noting that 2 2 = gG(z1) along BL+, the slope of the
vector field s2 is given by:
Proposition 1 Consider system (8). If(1 + f l ) < 0, then
as+ = ac+ and as- = ac-.
4683
The vector field points into non-saturated region L if sz > exists and has the same sign almost everywhere in F, then
s1 for s1 < 0 (fz < 0) and s2 < s1 for SI > 0 (fz > 0). F contains no closed trajectories.
Since both s1 and sz have fz in the denominator, the two For the non-saturated region without the boundaries ( L\
cases with fz > 0 and fi < 0 can be unified to give the (aL+UaL-)),using the condition (16), it can be seen that
following condition:
-1 - f1.1+
Xz fZ(1 - 1.1) < -f1+ Sgn(Z1) +
V h ( r ) = X1(1+ fi) A d 1 + fi(1 - 1211))
A1 (1 - 1.11)(1 +.I) (1 - 1.11)2 -Xlfzrzsgn(ii) < 0 (19)
Since (1 + X I ) t (1 - Ial) > 0 and (-fi sgn(z1)) + >
For the compact saturated region (D\ L )
(-f1 - l),the condition can be rearranged to give:

xl(i + fl)+ xZ(i + fz(l - i m - x ~ < o~ (13)~ z ~ V h ( z )= X I + Xz (20)

So, the lemma is proved if it can be shown that (13) is
verified. Case 1: X1 + Xz > 0
Case 1: 21 ? 0 In this case, the non-saturated region L is invariant (Lemma
From the Hurwitz stability condition, X1(1+ f l ) X z ( l + +
2). Since V h < 0 in the invariant region and there are no
f z ( l - 1.1 ) < 0. Since Xz < 0 and fi < 0 the last term limit cycles.
of (13), - X z f i q I O . So, the inequality (13) holds. Case 2: XI + Xz 5 0
Case 2: z1 < 0 In this case, V h is non-positive in both the saturated and
Substituting 11.1 = -21 in (13) reads: non-saturated regions, and so there are no limit cycles either.
The proof of the theorem is now based on the application
q5(21)=X1(1+fi)+Xz(l+fz)+Xz21(f~-fi) <0(14)
of the Poincare-Bendixsons theorem within the compact
Due to linearity, the maximum of q5(z1)V - 1 5 z1 5 0 invariant set D.So, every trajectory of (17) starting at zoE
occurs either at 21 = -1 or 11 = 0 depending on the sign C will either i) go to the equilibrium point at the origin, ii)
of (fz - fi). If (fz - f i ) I 0, the maximum is at 5 1 = 0. tend to a limit cycle or iii) be a limit cycle itself. In the first
+
Thus, it is required to prove Xl(l+ f i ) & ( 1 + fz) < 0, part of the proof, it has been shown that there are no limit
which is verified since it is the H w i t z stability condition. cycles. So, all trajectories in C converge asymptotically to
~ f l ) > 0, the maximum is at x1 = -1, and the the origin. Thus, the region of attraction is in fact A = C.
If ( f -
inequality (14) becomes:
(XI +X Z N + fd < 0.
Due to the Hurwitz stability condition, (l+fi)< 0, and due v. COMPARISON OF CONTROLLERS VIA SIMULATIONS
+
to the hypothesis, (XI Xz) > 0, this condition is verified.
System (1) is considered with the parameters A 1 = 1 and
The next theorem is the main result where the global Xz = -0.5, i.e.
asymptotic stability in C is guaranteed under the condition
Xl(2 + f l ) + Xz(l+ fz) < 0. (15)
Since X i > 0, this is a slightly more restrictive condition
than the Hunvitz condition X1(1+ fi) Xz(l+ fz) < 0. + The following 4 controllers are compared.
Theorem 2 r f f satisfies the condifions 1) Cl: Linear saturated controller:
U = sat(flzl +
fzz2), f i = -6 and fz = -3
(1 + fi + fz) < 0, Xi(2 + fi) + X z ( l + fz) < 0, (13 2) C2: Linear saturated controller with only rl fed
then, the closed loop sysfem back U = sat(flzl), f i = -3
3) C3: Time optimal controller 171:
5 = h ( z )= Ax + b sat(f1zl + fm(l- 1.11)) (17) Switching between U = -1 and U = 1
4) C4: Continuous nonlinear controller:
is asymptotically stable for all initial conditions in C.
, U = sat(flz1 + fizz(1 - Irll)),f 1 = -6
Proof First it is shown, using an extension of the and fz = -3
Bendixsons theorem 121, that no limit cycle exists. The The controllers C1 and C4 are tuned such that both closed-
extension of the Bendixsons theorem claims that if on a loop poles of the linear system are place at -2. As far as C2
compact invariant set F, the divergence V h defined by: is concerned, only the pole corresponding to the unstable
mode can be influenced by the controller and this pole is
placed at - 2 . The switching control law for C3 is given by
4684
the following set of equations: However, the damping of the stable state takes a long
time. The best time performance for both states is given
u.=sgn([ -1 I].)
by the time-optimal controller C3. The controller C4 needs
5, = x + uL,x,+= [ fXOZ0' ] more time to damp the states but does better than C1.
For initial conditions that are located nearer the origin, the
performance of C4 and C1 is similar.

The three properties mentioned in the introduction, i.e. (i)

global stability, (ii) local performance, and (iii) chattering
behavior are analysed for these four controllers.
A. Global sfabiliry
In Figure 2, the evolution of the closed-loop trajectories
T..
for.the initial condition xo = [ -0.8 2.8 ] is illustrated.
This initial condition is not in the region of attraction A of
C1 and therefore the corresponding trajectoy leaves the
null controllable region C and escapes to infinity.However,
all the other controllers are globally stable with A = C and Fig. 3. Phase diagram. The lines C+ and C- are the borders of the
non-saturated region of controller C4.
thus the corresponding trajectories converge to the origin.

I ,
.2 . ! %. '. ! .
! s. 1. !
! ,.c. !
.'
!
! !
I , , , , , , ,3. 5. , , ,
-3 4 0 as 4, 4.2 D 0.2 (I.. 0.1 0.1 I

X'

Fig. 4. Time evolution of 11 and 21

Fig. 2. Phase diagram. The lines C+ and C- are the borders of ihe
non-saturated region of conhuller C4.

C. Chaffering in the conhol signal

B. Local performance For this example, pseudo-random white noise perturba-
In Figure 3, the evolution of the trajectories of all tion is added to the measurements of the state variables.
T
controllers for the initial condition xo = [ 0.7 2.8 ] E C Figure 5 shows the control signal U for all controllers.
is shown. It is seen that the trajectory of controller C2 first It is seen that the,control signals of C1, C2 and C4
goes towards the x 2 axis since only the unstable state zI show no chattering behavior since they are generated by
is fed back. Then, convergence to the origin is ensured via continuous controllers. However, the control signal of C3
the open-loop dynamics of the stable state. shows chattering when the state is near the origin (after
t = 3.9[s]). Chattering is also present during the switching
The time evolution of the two states is shown in Figure from U = -1 to U = 1 at t = 1.826 , , . 1.833 [SI (Figure
4. For controller C2, the unstable state is damped quickly. 6).
4685
 I I
I I
t 3 6 3 7 4 4 , 2000.
Fig. 5 . Time evolution of the input U.
[2] M.S. Branicky. Multiple lyapunov functions and other analysis tools
for switched and hybrid systems. IEEE Tonsoclions on Aulomolic
Conirol, 43(4):475482, 1998.
[3] I-Y. Favez, E. Srinivasan, Ph. Mullhaupt, and D.Bonuin. Condition
for bifurcation of the region of attraction in linear planar systems with
saturated linear feedback. CDC 2002, 2002.
[4] T, Hu and 2. Lin. ConlmlSyslems wilh Acluolor Sarurarion: Anolyses
and De& Birkhauser, Boston, 2000.
[ 5 ] T. Hu, Z. Lin, and L. Qiu. Stabilization ofexpanentially unstable linear
systems with saturating achtators. IEEE Tronsocliom on Aulomalic
Contml, 46(6):973-979, 200 1.
[6] 2. Lin and A. Saberi. Semi-global exponential sIabilization of linear
systems subject to input saturation via linear feedback. System &
Control Leilers, 21:225-239, 1993.
[7] E.R. Pinch. Oplimol Conlml ond the Calculus of Voriolions. Oxford
University Press, New York, 1993.
[8] L. Scibile and B. Kouvaritakis. Stability region for a class of apen-
loop unstable linear systems: Theory and application. Aulomolico,
[9] H.J. SussmaM, E.D. Sontag, and Y. Yang. A general result on
the stabilization of linear systems wing bounded controls. IEEE
Tranroclions on Aulomolic donlmr, 39( li):2411-2424, 1994.

I 1.- . cm8 tm 1m ,134 ,as I.uI 1.-

I
,.w
1

Fig. 6. Time evolution of lht inpul U (zoomed: t = 1.826 . .. 1.833

[SI).

VI. CONCLUSION
In this paper, a simple continuous nonlinear controller for
the stabilisation of linear planar systems with one unstable
and saturated input was proposed. It was shown that this
controller is globally stabilising, i.e. its region of attraction
is the null controllable region. Furthermore, it satisfies
the desired performance locally around the origin. A third
propelty of this controller is the absence of chattering in
the control signal, which is a direct consequence of its
continuity.
Though this paper dealt only with planar systems with
one stable and one unstable pole, it is hoped that it can
be extended to systems with two unstable poles. Also, the
case with one unstable pole and multiple stable poles is of
interest.

REFERENCES
[I] I. Alvarer, R. Suirerer, and J. Alvarez. Planar linear systems with single
saturated feedback. Syslem & Confml Lefrers, 20319-326, 1993.
4686