Nonlinear To Linear System Project

Control of Nonlinear Systems
Nonlinear Control N. Marchand
Nicolas Marchand
References Outline Introduction
Linear/Nonlinear The X4 example
CNRS Associate researcher
Nicolas.Marchand@grenoble-inp.fr Control Systems Department, gipsa-lab

Grenoble, France
Linear approaches
Antiwindup Linearization Gain scheduling
Stability Nonlinear approaches

CLF Sliding mode Geometric control Recursive techniques X4 stabilization
Master PSPI 2009-2010

Nonlinear Control Master PSPI 2009-2010 1 / 174
Observers N. Marchand (gipsa-lab)
References
Nonlinear systems Khalil - Prentice-Hall, 2002
Nonlinear Control N. Marchand References Outline Introduction
Probably the best book to start with nonlinear control
Nonlinear systems S. Sastry - Springer Verlag, 1999

Good general book, a bit harder than Khalils
Mathematical Control Theory - E.D. Sontag - Springer, 1998

Mathematically oriented,
Can be downloaded at: http: // www. math. rutgers. edu/ ~ sontag/ FTP_ DIR/ sontag_ mathematical_ control_ theory_ springer98. pdf
Constructive nonlinear control - Sepulchre et al. - Springer, 1997 More

focused on passivity and recursive approaches
Linear approaches
Nonlinear control systems - A. Isidori - Springer Verlag, 1995

A reference for geometric approach
Applied Nonlinear control - J.J. Slotine and W. Li - Prentice-Hall, 1991

An interesting reference in particular for sliding mode

Research on Gain Scheduling - W.J. Rugh and J.S. Shamma Automatica, 36 (10):14011425, 2000 Survey of gain scheduling analysis and design - D.J. Leith and W.E. Leithead - Int. Journal of Control, 73:10011025, 1999
Surveys on Gain Scheduling,
The second reference can be downloaded at: http: // citeseer. ist. psu. edu/ leith99survey. html
Observers N. Marchand (gipsa-lab) Nonlinear Control Master PSPI 2009-2010 2 / 174
Outline
1
Introduction Linear versus nonlinear The X4 example Linear control methods for nonlinear systems Antiwindup Linearization Gain scheduling Stability Nonlinear control methods Control Lyapunov functions Sliding mode control State and output linearization Backstepping and feedforwarding Stabilization of the X4 at a position Observers
Linear approaches
3 4

Observers
N. Marchand (gipsa-lab)
Outline
1
Linear approaches
3 4

Observers
Introduction
Some preliminary vocable and denitions What is control ? A very short review of the linear case (properties, design of control laws, etc.) Why nonlinear control ? Formulation of a nonlinear control problem (model, representation, closed-loop stability, etc.)? Some strange possible behaviors of nonlinear systems Example : The X4 helicopter
Linear approaches

Outline
1
Linear approaches
3 4

Observers
Some preliminary vocable

We consider a dynamical system:

System u(t) x(t) y (t)
Linear approaches
where:
y is the output: represents what is visible from outside the system x is the state of the system: characterizes the state of the system u is the control input: makes the system move

The 4 steps to control a system

yd
k (x )
Controller
x ^
Observer
1
Linear approaches
Modelization
To get a mathematical representation of the system Dierent kind of model are useful. Often:
a simple model to build the control law a sharp model to check the control law and the observer

Design the state reconstruction: in order to reconstruct the variables needed for control Design the control and test it Close the loop on the real system
3 4
Observers
Linear dynamical systems

Some properties of linear system (1/2)
Denition: Systems such that if y1 and y2 are the outputs corresponding to u1 and u2 , then R: y1 + y2 is the output corresponding to u1 + u2 Representation near the operating point:
Transfer function: y (s ) = h(s )u (s ) State space representation: = Ax + Bu x y = Cx (+Du )
Linear approaches

The model can be obtained

physical modelization (eventually coupled with identication) identication (black box approach)
both give h(s ) or (A, B , C , D ) and hence the model.

Linear dynamical systems

Some properties of linear system (2/2)
= Ax + Bu x y = Cx (+Du ) Nice properties:

Unique and constant equilibrium Controllability (resp. observability) directly given by rank(B , AB , . . . , An1 B ) (resp. rank(C , CA, . . . , CAn1 )) Stability directly given by the poles of h(s ) or the eigenvalues of A (asymptotically stable < 0) Local properties = global properties (like stability, stabilizability, etc.) The time behavior is independent of the initial condition Frequency analysis is easy Control is easy: simply take u = Kx with K such that = Ax + BKx will (eig(A + BK )) < 0, the closed-loop system x be asymptotically stable
Linear approaches

Mathematical tool: linear algebra This is a caricature of the reality (of course problems due to uncertainties, delays, noise, etc.)
Nonlinear Control
Master PSPI 2009-2010 10 / 174
Why nonlinear control ?

Why nonlinear control if linear control is so easy ?
All physical systems are nonlinear because of

Actuators saturations Viscosity (proportional to speed2 ) Sine or cosine functions in robotics Chemical kinetic in exp(temperature) Friction or hysteresis phenomena ...
Linear approaches
More and more the performance specication requires nonlinear control (eg. automotive) More and more controlled systems are deeply nonlinear (eg. -nano systems where hysteresis phenomena, friction, discontinuous behavior) Nonlinear control is sometimes necessary (oscillators, cyclic systems, . . . )

Nonlinear dynamical systems

How to get a model ?
Representation:
State space representation:
ODE Ordinary dierential equation:
In this course, only:

(t ) x y (t ) = = f (x (t ), u (t )) h(x (t ), u (t ))
Linear approaches
PDE Partial dierential equation (trac, ow, etc.): 0 0 = =

x ,... ) (t ), f ( g (x (t ), x u (t )) x h(y (t ), x (t ), u (t ))

. . . Algebraic dierential equations (implicit), hybrid (with discrete or event based equations), etc.
The model can be obtained

physical modelization and then nonlinear identication of the parameters (identiability problems)

Open-loop control versus closed-loop control ?
yd
k (x ) Controller
u
States: x
x ^
Observer
Linear approaches
Open-loop control
nd u (t ) such that limt y (t ) yd (t ) = 0 Widely used for path planning problems (robotics)

Closed-loop control
nd u (x ) such that limt y (t ) yd (t ) = 0 Better because closed loop control can stabilize systems and is robust w.r.t. perturbation.
In this course
Only closed-loop control problems are treated
Observers

Aim of control ?
xd
k (x )
u
States: x

Controller
x ^
Observer
Tracking problem
nd u (x ) such that limt x (t ) xd (t ) = 0
Linear approaches
Stabilization problem
nd u (x ) such that limt x (t ) xd (t ) = 0 for xd (t ) = constant

Null stabilization problem

nd u (x ) such that limt x (t ) xd (t ) = 0 for xd (t ) = 0 In any cases, a null stabilization problem of z (t ) = y (t ) yd (t )
In this course
Only the stabilization problem will be treated.
Some strange behaviors of nonlinear systems

The undersea vehicle
Simplied model of the undersea vehicle (in one direction): (t ) = v |v | + u v Step answer:
1.5
Linear approaches
u (s ) = u (s ) =
2 s 1 s

No proportionality between the input and the output
0.5
u (s ) =
0 0 1 2 3 4 5 6 7 8 9 10
1 100s

The Van der Pol oscillator (1/2)
Van der Pol oscillator 1 = x2 x 2 = x1 h(x1 )x2 x

2 with h(x1 ) = 1 + x1
Oscillations: tunes the limit cycle

6 5 4 3 2 6
Linear approaches
= 1, x (0) = (5, 5) = 0.1, x (0) = (5, 5)
5 4 3 2
= 1, x (0) = (5, 5) = 0.1, x (0) = (5, 5)

1 0 1 2 3
x2
1 0 1 2 3
10
20
time
30
40
50
4 4
x1

The Van der Pol oscillator (2/2)
6 5 4 3 2
Oscillations: stable limit cycles, fast dynamics

6
= 1, x (0) = (0, 1) = 1, x (0) = (5, 5)
5 4 3 2
= 1, x (0) = (0, 1) = 1, x (0) = (5, 5)
1 0 1 2 3 6 5 30
x2
1 0 1 2 3 3
Linear approaches
10
20
time 4
3 2
40
50
x1

1 0 1 2 3
= 1, x (0) = (0, 1) = 1, x (0) = (5, 5)

0 0.1 0.2
time
0.3
0.4
0.5
Some strange behaviors of nonlinear systems 2. Representation of non linear systems

The tunnel diode (1/2)
The tunnel diode model dVC + iR dt
Example: the tunnel-d
= iL diL dt
E RiL = VC + L iR It gives: = h(vR )
Linear approaches

1 (t ) = x 2 (t ) = x
1 (h(x1 ) + x2 ) C 1 (x1 Rx2 + u ) L
with x1 = vC , x2 = iL and u = E

The tunnel diode (2/2)
1.2 1 1.4
The tunnel diode behavior: bifurcation with: 2 + 229.62v 3 226.31v 4 + 83.72v 5 iR = 17.76vR 103.79vR R R R
1.3
x (0) = (0, 1.23) x (0) = (0, 1.24)
1.2 1.1 1 0.9
x (0) = (0, 1.23) x (0) = (0, 1.24)
Linear approaches
0.8
0.6
x2
0.8 0.7 0.6 0.5
0.4

0.2
0.4 0 0 2 4 6 8 10 0 0.2 0.4 0.6 0.8 1
time
x1

The car parking
The car parking simplied model 1 = u1 x 2 = u2 x 3 = x2 u1 x System with a misleading simplicity
Linear approaches
Theorem (Brockett (83))

There is no (u1 (x ), u2 (x )) continuous w.r.t. x such that x is asymptotically stable

manoeuvres discontinuous control time-varying control
Marchand and Alamir (2003)
In practice:

Everything is possible
= f (x , u ) x y = h (x , u ) Everything is possible:
Equilibrium can be unique, multiple, innite or even not exist Controllability (resp. observability) are very hard to prove (it is often even not checked) Stability may be hard to prove Local properties = global properties (like stability, stabilizability, etc.) The time behavior is depends upon the initial condition Frequency analysis is almost impossible No systematic approach for building a control law: to each problem corresponds it unique solution
Linear approaches

Mathematical tool: Lyapunov and dierential tools
Outline
1
Linear approaches
3 4

Observers
The X4 helicopter
How it works ?
Principe de fonctionnement 4 xed rotors with controlled rotation speed si 4 moteurs

s3 s4
Linear approaches
s2 s1

The X4 helicopter
How it works ?
Principe de fonctionnement 4 xed rotors with controlled rotation speed si 4 moteurs 4 generated forces prop. Fi 4 forces
F3 F4
F2 F1
Linear approaches

The X4 helicopter
How it works ?
Principe de fonctionnement 4 xed rotors with controlled rotation speed si 4 moteurs 4 generated forces prop. Fi 4 forces 4 counter-rotating 4 couplestorques prop. i
3 4 2 1
Linear approaches

The X4 helicopter
How it works ?
Linear approaches
Principe de fonctionnement 4 xed rotors with controlled rotation speed si 4 moteurs 4 generated forces prop. Fi 4 forces 4 counter-rotating 4 couplestorques prop. i Contrle : generated with a Roll movement Roulis : dissymmetry between left and right forces: F2 1 r = l (F4 F2 ) l )
F4

The X4 helicopter
Linear approaches

Principe de fonctionnement 4 xed rotors with controlled rotation speed si 4 moteurs 4 generated forces prop. Fi 4 forces F3 4 counter-rotating 4 couplestorques prop. i Contrle : generated with a Roll movement Roulis : dissymmetry between left and right Tangage : forces: 1 r = l (F4 F2 ) l ) Pitch movement generated with a F1 dissymmetry between front and rear forces: p = l ( F1 F3 )
How it works ?
The X4 helicopter
How it works ?
Linear approaches

Principe de fonctionnement 4 xed rotors with controlled rotation speed si 4 moteurs 4 generated forces prop. Fi 4 forces 4 counter-rotating 4 couplestorques prop. i Contrle : generated with a Roll movement Roulis : dissymmetry between left and right Tangage : 3 forces: Lacet : r = l (F4 F2 ) Pitch movement generated with a dissymmetry between front and rear 2 forces: p = l ( F1 F3 ) Yaw movement generated with a dissymmetry between front/rear and left/right torques: y = 1 + 3 2 4
Nonlinear Control
Observers N. Marchand (gipsa-lab) Master PSPI 2009-2010 23 / 174
The X4 helicopter
Building a model (1/3)
Electrical motor: A 2nd order system with friction and saturation usually approximated by a 1rst order system: i s =
2 km 1 km si load + satU i (Ui ) Jr R Jr Jr R
i {1, 2, 3, 4} (1)
Linear approaches
si : rotation speed Ui : voltage applied to the motor; real control variable load : motor load: load = kgearbox |si | si with drag coecient
Aerodynamical forces and torques: Very complex models exist

The X4 helicopter
Electrical motor: A 2nd order system with friction and saturation usually approximated by a 1rst order system: i s =
i {1, 2, 3, 4} (1)
Linear approaches
si : rotation speed Ui : voltage applied to the motor; real control variable load : motor load: load = kgearbox |si | si with drag coecient
Aerodynamical forces and torques: Very complex models exist but overcomplicated for control, better use the simplied model: Fi r p y = = = = bsi2 2 s 2) lb (s4 2 2 s 2) lb (s1 3 2 + s 2 s 2 s 2) (s1 3 2 4

i {1, 2, 3, 4}
(2)
b : thrust coecient, : drag coecient

The X4 helicopter
t3 Two frames
a xed frame E (e1 , e2 , e3 ) a frame attached to the X4 T (t1 , t2 , t3 )
t2 t1 e2 e1
Frame change
a rotation matrix R from T to E
e3
Linear approaches
State variables:
Cartesian coordinates (in E )
position p velocity v

Attitude coordinates:
angular velocity in the moving frame T either: Euler angles three successive rotations about t3 , t1 and t3 of angles angles , and giving R or: Quaternion representation (q0 , q ) = (cos /2, u sin /2) represent a rotation of angle about u
The X4 helicopter
Cartesian coordinates: = v p = mg e3 + R ( mv Attitude:

Euler angles formalism: R J = = R J + tot
0 with = 3 2 3 0 1 2 1 0
Fi (si )t3 )
(3)
Linear approaches
(4)

is the skew symmetric tensor associated to Quaternion formalism: = 1 ()q q 2 () = = 1 ( q ) with 2 (q ) = J = J + tot
T q T I33 q0 + q 0
(5)
where tot =
i
Ir t3 si +pert
gyroscopic torque
Nonlinear Control
r (s2 , s4 ) + p ( s 1 , s 3 ) y (s1 , s2 , s3 , s4 )
Master PSPI 2009-2010 26 / 174
The X4 helicopter
Review of the nonlinearities
Linear approaches

2 kgearbox i = Jkrm s |si | si + Jkrm i (Ui ) R si Jr R satU p = v 0 = mg e3 + R mv 0 i F i ( si ) R = R 0 r ( s 2 , s 4 ) 0 + = J p (s1 , s3 ) J i Ir y ( s 1 , s 2 , s 3 , s 4 ) i si In red: the nonlinearities In blue: where the control variables act
The X4 helicopter
Identication of the parameters
Electrical motor:
For small very close to a linear Winter 2004-2005 CoaX input steps, the system behaves rst order system Lorsquon bouge laxe, lhlice rpond la perturbation dune manire similaire aux hlices Hence, use linear identication tools Proxflyer ce qui nous promet de bons rsultats sur le CoaX. Malheureusement nous navons pas i is found U on the data-sheet ofsur the motorcontrarotatif (damage avoidance) pu tester le comportement du systme avec les deux hlices le systme nayant
pas deux hlices disposition avant le montage final.
Autonomous Systems Laboratory
Aerodynamical parameters: b and
b and measured with specic test beds, 5.5.2 Mcanisme de contra rotation avec hlices blanches
Nous avons fait tourner le module muni dhlices blanches sur une balance pour avoir quelques chiffres quant la qualit du systme. Les hlices blanches ne sont pas les plus optimise pour notre application, mais elle donne dj des informations utiles pour la suite du projet. Les rsultats sont reprsents dans un tableau et des graphiques plus loin.
Linear approaches

Figure 36 : Hlices blanches
Figure 35 : Ban de test utilis
Observers
Pour tester les hlices ou les moteurs, nous utilisons, en temps normal, un ban de test spcialement conu cet effet. La force de pousse est transmise la balance via un levier (voir N. Marchand (gipsa-lab ) Nonlinear Control Master PSPI 2009-2010 28 / 174
The X4 helicopter
Identication of the parameters
Electrical motor:
For small input steps, the system behaves very close to a linear rst order system Hence, use linear identication tools i is found on the data-sheet of the motor (damage avoidance) U
Aerodynamical parameters: b and

b and measured with specic test beds, depends upon temperature, distance from ground, etc.
Linear approaches
Mechanical parameters:
l m J Ir length of an arm of the helicopter, easy to measure total mass of the helicopter, easy to measure body inertia, hard to have precisely rotor inertia, hard to have precisely

The X4 helicopter
Values of the parameters
Motor parameters:
parameter km Jr R kgearbox i U description motor constant rotor inertia motor resistance gearbox ratio maximal voltage value 4.3 103 3.4 105 0.67 2.7 103 12 unit N.m/A J.g.m2 V
Aerodynamical parameters:
parameter b description thrust coecient drag coecient value 3.8 106 2.9 105
Linear approaches

Body parameters:
parameter J m l g description inertia matrix mass of the UAV radius of the UAV gravity value 14.6 103 0 0 0 7.8 103 0 0 0 7.8 103 0.458 22.5 9.81 unit kg.m2 kg cm m/s2
The X4 helicopter
Open-loop behavior with roll initial speed
Linear approaches

The X4 helicopter
Open-loop behavior with pitch initial speed
Linear approaches

The X4 helicopter
Open-loop behavior with yaw initial speed
Linear approaches

The X4 helicopter
Open-loop behavior with roll and yaw initial speed
Linear approaches

The X4 helicopter
Open-loop behavior with pitch and yaw initial speed
Linear approaches

Outline
1
Linear approaches
3 4

Observers
Outline
1
Linear approaches
3 4

Observers
Linear systems with saturated inputs

Impact of input saturations on the control of the X4s rotors (1/4)
We go back to the X4 example and focus on the rotors: i = s

Linear approaches
If one wants to act on the X4 with desired forces Fid , it is necessary to be able to set the rotors speeds si to sid with sid = 1 d F b i

A usual way to control the electrical motor consist in

taking load as un unknown load i neglecting the voltage limitations U

The so obtained system is linear si (s ) Ui (s ) =

1 km R + Jkr 2 s m
G 1 + s
Dene a PI controller for it: C (s ) = Kp + Taking Ki = is:

1 CL G
Linear approaches
Ki s

and Kp = Ki , the closed loop system si ( s ) Ui (s ) 1 1 + CL s

600
Make a step that compensates the weight, that is such d that sid = mg i Fi = mg 4b so that Taking CL = 50 ms, one gets with saturations
400
Linear approaches
si
200

80 60 40
Ui
20 0 20 0 1 2 3 4 5
Observers N. Marchand (gipsa-lab) Nonlinear Control
time
Master PSPI 2009-2010 39 / 174

Saturation may cause instability
Pulse Generator
The result could be worse:

7s+5 s Control Saturation 1 s-1 Transfer Fcn
For u [1.2, 1.2], the closed-loop behavior is:

4 3.5 3 2.5 2 1.5 1
Linear approaches
with saturation without saturation

0.5 0 0.5 1 0 1 2 3 4 5
Saturations may lead to instability especially in the presence of integrators in the loop

6. Linear systems with a nonlinearity
Key idea of the anti-windup scheme

Consider a linear system with a PID controller:

PID controller Linear system
Linear approaches
Kd Kp KI

1/s
The instability comes from the integration of the error Key idea: soften the integral eect when the control is saturated

6. Linear systems with a nonlinearity
PID controller with anti-windup

Structure of the PID controller with anti-windup:

Anti Windup PID controller Linear system
Kd Kp KI
Linear approaches
1/s Ks

, that is if u is not saturated, then the PID If u = u controller with anti-windup is identical to the classical PID controller ), Ks tunes the reduction of the If u is saturated (u = u integral eect of the PID

General controller with anti-windup
Structure of the general dynamic controller: D F e G

+ +
Linear approaches
1 s
+ +

with dynamics given by: c x u = Fxc + Ge = Cxc + De

General controller with anti-windup
Anti-windup structure of the general dynamic controller: D

F KC
e
G KD
+ +
1 s
C K
+ +
Linear approaches
with dynamics given by: c x = Fxc + Ge + K (u Cxc De ) = (F KC )xc + (G KD )e + Ku

Choice of the antiwindup parameters

Take K such that (F KC ) is stable and |max (F KC )| > |max (F )|

A short Lyapunov explanation of the behavior
Consider a stable closed loop linear system: it is globally asymptotically stable. A saturation on the input may:
transform the global stability into local stability. In this case, the aim of the anti-windup is to increase the radius of attraction of the closed loop system keep the global stability property. In this case, the aim of the anti-windup is to renders the saturated system closer to its unsaturated equivalent
Linear approaches

However, there is no formal proof of stability of anti-windup strategies
Other approaches for saturated inputs

Optimal control Nested saturation function (under development)

Back to the unstable case
Pulse Generator
The unstable case:

7s+5 s Control Saturation 1 s-1 Transfer Fcn
The closed-loop behavior is:

4 3.5 3 2.5 2
Linear approaches
with saturation with saturation and anti-windup
without saturation
1 0.5 0 0.5 1 0 1 2
1.5

However, nothing is magic and divergent behavior may occur because of the level of the step input

600
Make a step that compensates the weight, that is such d that sid = mg i Fi = mg 4b so that Taking CL = 50 ms, one gets with anti-windup
400
Linear approaches
si
200

80 60 40
Ui
20 0 20 0 1 2 3 4 5
time
Master PSPI 2009-2010 47 / 174

Linear approaches
3 4

Linearization of nonlinear systems

Impact of the load on the control of the X4s rotors
Nonlinear Control N. Marchand References
600
Make a step that compensates the weight, that is such that sid = that = mg Taking CL = 50 ms, one gets with load
d i Fi
mg 4b
so
Outline Introduction
400
si
200
Linear approaches
20 10 0

Ui
10 20
time
The PI controller seems badly tuned: for t > 1.3 s , the control is not saturated but the convergence is very slow. Whats wrong ?

Back to the rotor dynamical equation: i = s

2 kgearbox km km si |si | si + satU i (Ui ) Jr R Jr Jr R
Linear approaches
Taking load as unknown implies that the PI control law was tuned for: i = s
2 km km si + satU i (Ui ) Jr R Jr R

Implicitly, the second order terms were neglected Unfortunately, these two systems behave similarly i si is 1 small, which is not the case for sid = mg 4b 544 rad s

Linearization at the origin
Linearization at the origin

Take a nonlinear system x y = f (x , u ) = h (x , u )
Linear approaches

with f (0, 0) = 0. Then, near the origin, it can be approximated by its linear Taylor expansion at the rst order: f f = x x+ u x u (x =0,u=0) (x =0,u =0) y h(0, 0) =
A B
h x
h x+ u (x =0,u =0)
C
u
(x =0,u =0) D

Linearization at a point
Linearization at a point
Take a nonlinear system x y = f (x , u ) = h (x , u )
Linear approaches

then, near the equilibrium point (x0 , u0 ), it can be approximated by its linear Taylor expansion at the rst order: f f x = (x x0 ) + (u u0 ) x (x0 ,u0 ) u (x0 ,u0 ) x x u y h(x0 , u0 ) =
A B
h x
(x x0 ) +
(x0 ,u0 ) C x
h u
(u u0 )
(x0 ,u0 ) D u
= x x0 and u = u u0 which is linear in the variables x

Some properties of the linearization
Controllability properties
Take a nonlinear system = f (x , u ) x and its linearization at (x0 , u0 ): = Ax + Bu x (7) (6)
Linear approaches
If (7) is controllable then (6) is locally controllable Nothing can be concluded if (7) is uncontrollable

Example
The car is controllable but not its linearization 1 = u1 1 x x 1 0 x 2 = u2 2 = 0 1 u1 u2 x linearization at the origin 3 = x2 u1 3 x x 0 0
Observers

Some properties of the linearization
Stability properties
Take a nonlinear system = f (x , u ) x and its linearization at (x0 , u0 ): = Ax + Bu x = k (x ) so that: Assume that one can build a feedback law u (9) is asymptotically stable ( < 0): then (8) is locally ) asymptotically stable with u = u0 + k (x (9) is unstable ( > 0): then (8) is locally unstable with ) u = u0 + k (x Nothing can be concluded if (9) is simply stable ( 0) (9) (8)
Linear approaches


Back to the rotor dynamical equation: i = s

2 kgearbox km km si |si | si + satU i (Ui ) Jr R Jr Jr R
Dene Uid , the constant control that keeps the steady state speed: Uid = km sid + Rkgearbox d d si si km
Linear approaches
The linearization of the rotor dynamics near sid gives: i = s with:

2 2kgearbox d km + si Jr R Jr

si +
km i ) satU i (U Jr R
i = Ui U d U i si = si sid i ) = U sat i (U
i U i U i U
i , +U i ] if Ui [U i if Ui U i if Ui U
Master PSPI 2009-2010 55 / 174
Nonlinear Control

600
Make a step that compensates the weight, that is such that d sid = mg i Fi = mg 4b so that Taking CL = 50 ms and a PI controller tuned at sid on the system with load, one has:
400
Linear approaches
si
200

20 10 0
Ui
10 20 0 1 2 3 4 5
time

Linear approaches
3 4

Towards gain scheduling

Impact of a change in the steady state speed of the X4s rotors
400
Take again CL = 50 ms and a PI controller tuned at sid Make speed steps of dierent level
600
200
si
0 200
Linear approaches
10
20 10 0

Ui
10 20 0 2 4 6 8 10
time
The controller is well tuned near sid but not very good a large range of use

Local linearization may be inadequate for large excursion
The result could be worse: Take the inverted pendulum
1 ()
m+M ml cos ml cos I + ml 2
mgl sin 2 sin k y u + ml
where
I is the inertia of the pendulum () = (I + ml 2 )(m + M ) m2 l 2 cos2
Linear approaches

Linearize it near the upper position y, y ): k = 0 and x = (, , 1 0 1 0 x 2 x 2 = 3 0 0 x 3 0 0 0 4 x 1 3 0 0

Nonlinear Control
( = 0) with m = M = I = g = 1, 0 0 x1 x2 1 0 + 3 u 1 x3 0 2 x4 0 3
Master PSPI 2009-2010 59 / 174

Local linearization may be inadequate for large excursion
Build a linear feedback law that places the poles at -1 Starting from (0) = /5, it converges, from (0) = 1.1 /5, it diverges
15 15
Inverted pendulum
Linearization of the inverted pendulum Initial condition: x (0) = (/5, 0, 0, 0)
Linear approaches
10
10
x (0) = (1.1/5, 0, 0, 0)

10
15
10
15

For some systems, the radius of attraction of a stabilized linearization may be very small (less than one degree on the double inverted pendulum for instance) Linearization is not suitable for large range of use of the closed-loop system For other systems, the controller must be very nely tuned in order to meet performance requirements (e.g. automotive with more and more restrictive pollution standards) Linearization is not suitable for high accuracy closed loop systems Either for stability reasons or performance reasons, it may be necessary to tune the controller at more than one operating point
Linear approaches


Gain scheduling uses the idea of using a collection of linearization of a nonlinear system at dierent operating points
Linear approaches

Gain scheduling
Take a nonlinear system
= f (x , u )
Dene a family of equilibrium points s = (xeq , ueq ) such that f (xeq , ueq ) = 0
Linear approaches
At each point s , linearize the system assuming s is constant: = A(s )x + B (s )u x = x xeq and u = u ueq with x At each point s , dene a feedback law: u = ueq (s ) K (s )(x xeq (s )) with (Eig(A(s ) B (s )K (s ))) < 0

Gain scheduling
The controller is then obtained by changing s in the a priori dened collection of points The change of s can be done discontinuously or continuously by interpolation
Linear approaches

Gain scheduling
Drawbacks of gain scheduling :
Convergence only if s varies slowly The performance within a linearization area may be poor The mapping may be prohibitive if the number of states is large The stability issues are not clear
Recent evolution of gain scheduling:

Dynamic gain scheduling: improve the transition between the dierent controllers, limits the importance of the slow motion of s LPV: Linear parameter varying methods considers the nonlinear system as x = A((x ))x + B ((x ))u
Linear approaches

and, under some conditions, uses LMI to build a feedback. Improves the performance within a linearization area.
Gain scheduling
Handling changes in the steady state speed of the X4s rotors
400
Take again CL = 50 ms and a PI controller tuned at si Make speed steps of dierent level
600
Outline
200
Introduction
si
0 200
Linear approaches
20 10 0
10

Ui
10 20 0 2 4 6 8 10
time
The rotors are now well controlled
Stability
1
Linear approaches
3 4

Observers
Stability
Consider the autonomous nonlinear system:
= f (x , u (x )) = g (x )
(10)
Stability
System (10) is said to be stable at the origin i: R > 0, r (R ) > 0 such that x0 B (r (R )), x (t ; x0 ), solution of (10) with x0 as initial condition, remains in B (R ) for all t > 0.
Linear approaches
Attractivity
The origin is said to be attractive i: limt x (t ; x0 ) = 0.

Asymptotic stability
System (10) is said to be asymptotically stable at the origin i it is stable and attractive
Observers
Stability
Graphical interpretation
Linear approaches

A For nonlinear systems: Attractivity

Nonlinear Control
For linear systems: Attractivity Stability Stability Stability and attractivity : properties hard to check ?
Observers N. Marchand (gipsa-lab) Master PSPI 2009-2010 69 / 174
Stability
Using the linearization
Asymptotic stability and local linearizaton

= g (x ) and its linearization at the origin Consider x g = x x . Then: x
x =0
Linear approaches
Linearization with eig< 0 asymptotically stable Linearization with eig> 0 unstable
Nonlinear system is locally Nonlinear system is locally

Linearization with eig= 0: nothing can be concluded on the nonlinear system (may be stable or unstable) Only local conclusions
Stability
Lyapunov theory: Lyapunov functions
Aleksandr Mikhailovich Lyapunov

Markovs school friend, Chebyshevs student
Master Thesis : On the stability of ellipsoidal forms of equilibrium of a rotating liquid in 1884 Phd Thesis : The general problem of the stability of motion in 1892
6 June 1857 - 3 Nov 1918
Linear approaches
Denition: Lyapunov function

Let V : Rn R be a continuous function such that:
1

(denite) V (x ) = 0 x = 0 (positive) x , V (x ) 0 (radially unbounded) lim

x V (x )
2 3
= +
Lyapunov functions are energies


Stability
June Lyapunov 1857 - 3 Nov 1918 Lyapunov 6 theory: theorem

Aleksandr Mikhailovich Lyapunov

6 June 1857 - 3 Nov 1918
Linear approaches
Theorem: (First) Lyapunov theorem

If a Lyapunov function V : Rn R+ s.t. (strictly decreasing) V (x (t )) is strictly decreasing for all x (0) = 0

then the origin is asymptotically stable. If a Lyapunov function V : Rn R+ s.t. (decreasing) V (x (t )) is decreasing then the origin is stable.
Observers
Stability
Lyapunov theorem: a graphical interpretation
Linear approaches

Stability
Stability and robustness
Stability holds robustness: = V g (x ) < 0 = g (x ) stable V x x V (g (x ) + (x )) < 0 x = g (x ) + (x ) stable x
Linear approaches

Outline
1
Linear approaches
3 4

Observers
The X4 helicopter
The control problems
Linear approaches
i s p mv R J
= = v = = =
2 km Jr R s i
kgearbox Jr
|si | si +
mg e3 + R R J

Attitude control problem 0 r (s2 , s4 ) 0 + p (s1 , s3 ) i Ir y ( s 1 , s 2 , s 3 , s 4 ) i si
0 0 i F i ( si )
Position control problem
km Jr R
satU i (Ui )
Outline
1
Linear approaches
3 4

Observers
Control Lyapunov functions

Characterizing property (1/3)
Control Lyapunov functions (CLF) where introduced in the 80s CLF are for controllability and control what Lyapunov functions are for stability Lyapunov: a tool for stability analysis of autonomous systems
Take 1 x 2 x Take the Lyapunov function V (x ) = 3 2 2 x + x1 x2 + x2 2 1 = = x2 x1 x2
Linear approaches

Along the trajectories of the system: (x (t )) V = V (x ).f (x ) = x

2
hence, V 0 and the system is asymptotically stable


Control Lyapunov Functions: a tool for stabilization of controlled systems

Take now the controlled system 1 x 2 x = = x2 x1 + u
Linear approaches
Take the Lyapunov function V (x ) = 3 2 2 x + x1 x2 + x2 2 1

Along the trajectories of the system: (x (t ), u (t )) V = V ( x ) .f ( x , u ) 2 2 = x1 + x1 x2 + x2 + (x1 + 2x2 )u

Control Lyapunov Functions: a tool for stabilization of controlled systems

Along the trajectories of the system: (x (t ), u (t )) V = V ( x ) .f ( x , u ) 2 2 = x1 + x1 x2 + x2 + (x1 + 2x2 )u
Linear approaches
hence
if x1 + 2x2 = 0, there exists u such that V otherwise (if x1 + 2x2 = 0) (x (t ), u (t )) V =
2 2 2 2 4x2 2x2 + x2 = 5x2

which is negative unless x2 = 0 = 1 x : V 2 1
In conclusion: characterizing property of CLF

(x (t ), u (t )) < 0 For any x = 0, there exists u such that V

Formal denitions
Denition: Lyapunov function

A continuous function V : Rn R0 is called a Lyapunov function if
1
[positive deniteness] V (x ) 0 for all x and V (0) = 0 if and only if x = 0 [radially unbounded] limx V (x ) = + [positive deniteness] V (x ) 0 for all x and V (0) = 0 if and only if x = 0 [proper] for each a 0, the set {x |V (x ) a} is compact (, K ) ( x ) V (x ) ( x ) x Rn
Linear approaches
or:
1

or, alternatively:

Formal denitions
Denition: Control Lyapunov Function

A dierentiable control Lyapunov function denotes a dierentiable Lyapunov function being innitesimally decreasing, meaning that there exists a positive continuous denite function W : Rn R0 and such that:
m x Rn u R
Linear approaches
sup min V (x )f (x , u ) + W (x ) 0

Roughly speaking a CLF is a Lyapunov function that one can force to decrease Extensions to nonsmooth CLF exist and are necessary for the class of system that can not be stabilized by means of smooth static feedback

Control ane systems
Denition: Directional Lie derivative

For any f and g with values in appropriate sets, the Lie derivative of g along f is dened by: Lf g (x ) := g (x ) f (x ) Iteratively, one denes:
k 1 Lk g (x ) f g (x ) := Lf Lf
Theorem
= g0 ( x ) + Consider a control ane system x f (0) = 0. Assume that the set
m i =1 ui gi (x )
Linear approaches
with f smooth and

S = x |Lf V (x ) = 0 and Lk f Lgi V (x ) = 0 for all k N, i {1, . . . , m} then, the feedback k (x ) := (V (x ) G (x ))T = Lg1 V (x ) globally asymptotically stabilizes the system at the origin. Lgm V (x )
T
= {0}

Control ane systems
Theorem: Sontags universal formula (1989)

= g0 ( x ) + m Consider a control ane system x i =1 ui gi (x ) with f smooth and f (0) = 0. Assume that there exists a dierentiable CLF, then the feedback ki (x ) = bi (x )(a(x ), (x )) (= 0 for x = 0)
Linear approaches

a(x ) := V (x )f (x ) and bi (x ) := V (x )gi (x ) for i = 1, . . . , m B (x ) = b1 (x ) b2 (x ) and (x ) = B (x ) 2 q : R R is any real analytic function with q (0) = 0 and bq (b ) > 0 for any b>0 is the real analytic function: a+ a2 +bq (b ) if b = 0 b ( a , b ) = 0 if b = 0 is such that k (0) = 0, k smooth on Rn \0 and
x Rn
sup V (x )f (x , k (x )) + W (x ) 0

Control ane systems
Denition: Small control property

A CLF V is said to satises the small control property if for any , there is some so that:
x B() u B()
sup
min V (x )f (x , u ) + W (x ) 0
Linear approaches
The small control property implicitly means that if is small (hence the control), the system can still be controlled as long as it is suciently close to the origin

Theorem: Sontags universal formula (1989)

= g0 (x ) + i = 1m ui gi (x ) Consider an control-ane system x with f smooth and f (0) = 0. Assume that there exists a dierentiable CLF, then the previous feedback k with q (b ) = b is smooth on Rn \0 and continuous at the origin

Robustness issues
Model = f (x , u ) x u = k (x )
Real system = f (x , u )+ x u = k (x +)
Denition: Robustness w.r.t. model errors

A feedback law k : Rn Rm is said to be robust with respect to model errors if limt x (t ) B (r ()) with lim0 r () = 0 (x (t ) denotes the solution of = f (x , k (x )) + ) x
Linear approaches
Denition: Robustness w.r.t. measurement errors

A feedback law k : Rn Rm is said to be robust with respect to measurement errors if limt x (t ) B (r ()) with lim0 r () = 0 (x (t ) denotes the solution = f (x , k (x + ))) of x

Theorem: Smooth CLF = robust stability (1999)

The existence of a feedback law robust to measurement errors and model errors is equivalent to the existence of a smooth CLF

Conclusion
A CLF is a theoretical tool, not a magical tool A CLF is not a constructive tool Finding a stabilizing control law and nding a CLF are at best the same problems. In all other cases, nding a stabilizing control law is easier than nding a CLF Even if one can nd a CLF for a system, the universal formula often gives a feedback with poor performances However, this is an important eld of research in the control system theory community that proved important theoretical results, for instance the equivalence between asymptotic controllability and stabilizability of nonlinear systems (1997)
Linear approaches

Other structural tools

Passivity
Denition: Passivity
A system = f (x , u ) x y = h(x , u )
is said to be passive if there exists some function S (x ) 0 with S (0) = 0 such that
T
Linear approaches
S (x (T )) S (x (0))
u T ()y ()d
0

Roughly speaking, S (x (0)) (called the storage function) denotes the largest amount of energy which can be extracted from the system given the initial condition x (0). Passivity eases the design of control law and is a powerful tool for interconnected systems Important literature
Other structural tools

Input-to-State Stability (ISS)
Class K A continuous function : R0 R0 is said in K if it is strictly increasing and (0) = 0 Class KL A continuous function : R0 R0 R0 is said in KL if (, s ) K for each xed s and (r , ) is decreasing for each xed r and lims (r , s ) = 0
Denition: Input-to-State Stability

A system = f (x , u ) x is said to be input-to-state stable if there exists functions KL and K such that for each bounded u () and each x (0), the solution x (t ) exists and is bounded by: x (t ) ( x (0) , t ) + ( sup u () )
0t
Linear approaches

Roughly speaking, ISS characterizes a relation between the norm of the state and the energy injected in the system: the system can not diverge in nite time with a bounded control ISS eases the design of control law Important literature
Outline
1
Linear approaches
3 4

Observers
Sliding mode control

An introducing example
Take again the linear system 1 = x2 x 2 = x1 + u x Assume u is forced to remain in [1, 1]

In practical applications, u is always constrained in an interval [u , u ]. We will see later on how to handle it. To begin, we take the above simplied constraint.
Linear approaches
Take the CLF 3 2 2 + x1 x2 + x2 = xT V (x ) = x1 2

3 2 1 2 1 2

Goal
Find u in [1, 1] that makes V decrease as fast as possible

: Making V decrease as fast as possible is equivalent to minimizing V u = Argmin V

u [1,1] 2 2 2 2 x1 = Argmin 1 x2 2 + x2 1 + 2x2 2 )u 1 + x1 2 + (x1 u [1,1]
= sign(x1 + 2x2 )
can not be changed
minimal for u = sign(x1 +2x2 )
Simulating the closed loop system with initial condition x (0) = (2, 3), it gives:
2
Linear approaches
x2 x1
0 2 4 6 8 10
Stability
1.5
Nonlinear approaches
1 0.5 0 0.5 1 1.5 0 2 4 6 8 10
Why does it work ?

Dene (x ) = x1 + 2x2 . The set {S (x ) = {x |(x ) = 0}} {|x1 | 0.6, |x2 | 0.3} is attractive:
S = {x
|x
2x
0}
Linear approaches

x1
x2

Dene (x ) = x1 + 2x2 . The set {S (x ) = {x |(x ) = 0}} {|x1 | 0.6, |x2 | 0.3} is attractive:
S = {x
|x
2x
0}
Linear approaches

<0 SS >0 SS
>0 SS <0 SS
x1
x2

The set {S (x ) = {x |(x ) = 0}} {|x1 | 0.6, |x2 | 0.3} is attractive Once the set is joined, the control is such that x remains on S (x ), that is: (x ) = 0 = x 1 + 2x 2 = x2 x1 + u S Hence, on S (x ), u = sign(x1 + 2x2 ) has the same inuence on the system as the control ueq = x1 x2 On S (x ), the system behaves like: 1 = x2 x 2 = x1 + u = x2 x x2 and hence also x1 clearly exponentially go to zero without leaving {S (x ) = {x |x1 + 2x2 = 0}} {|x1 | 0.6, |x2 | 0.3}
Linear approaches


Principle of sliding mode control
Principle: Dene a sliding surface S (x ) = {x |(x ) = 0} A stabilizing sliding mode control is a control law
discontinuous in on S (x ) that insures the attractivity of S (x ) such that, on the surface S (x ), the states slides to the origin
Linear approaches

Main characteristic of sliding mode control: Robust control law Discontinuities may damage actuators (ltered versions exist)

Notion of sliding surface and equivalent control
With a sliding mode control, the system takes the form: = x f + (x ) f (x ) if (x ) > 0 if (x ) < 0
f + (x ) (x ) > 0 fn+ (x ) fn (x ) f (x )
Linear approaches

(x ) = 0
(x ) < 0
What happens for (x ) = 0 ? = f + (x ) + (1 )f (x ) The solution on (x ) = 0 is the solution of x where satises fn+ (x ) + (1 )fn (x ) = 0 for the normal projections of f + and f

A sliding mode control for a class of ane systems
Theorem: Sliding mode control for a class of nonlinear systems

Take the nonlinear system: 1 f1 (x ) + g1 (x )u x x x1 2 = = f (x ) + g (x )u . . . . . . n xn1 x Choose the control law u= p T f (x ) T sign((x )) T p g (x ) p g (x )
Linear approaches

with > 0, (x ) = p T x and p T = p1 pn is a stable polynomial (all roots have strictly negative real parts) Then the origin of the closed loop system is asymptotically stable

Sliding mode control for some ane systems: proof of convergence
Proof: Take the Lyapunov function 1 V (x ) = x T pp T x 2

(x ) Note that V (x ) = 2 Along the trajectories of the system, one has:
2
Linear approaches
(x ) = T (x ) (x ) = x T p p T f (x ) + p T g (x )u V with the chosen control, it gives: (x ) = (x ) sign((x )) = |(x )| V x tends to S = {x |(x ) = 0} tunes how fast the system converges to S = {x |(x ) = 0}


Sliding mode control for some ane systems: proof of convergence
On S , the system is such that (x ) = 0 = p1 x1 + + pn1 xn1 + pn xn But: n xn1 = x n1 = xn xn2 = x . . .

(2)
Linear approaches
2 = = xn x1 = x That can be written with zi = xni +1 in the 0 1 0 0 0 1 . .. = z . . . 0 p3 2 p p1 p1
(n1)
form: .. . .. . 0 0 . . . 0 z 1 pn p1

Hence, x asymptotically goes to the origin

Nonlinear Control
Master PSPI 2009-2010 99 / 174

Sliding mode control for some ane systems: time to switch
Consider an initial point x0 such that (x0 ) > 0. Since (x ) = (x ) sign((x )) (x ) it follows that as long as (x ) > 0 (x ) = Hence, the instant of the rst switch is ts = (x0 ) <
Linear approaches

Moreover, ts 0 as

Robustness issues
Assume that one knows only a model ^ (x ) + g =f x ^ (x )u of the true system = f (x ) + g (x )u x The time derivative of the Lyapunov function is:
T ^g T )p pT g ^f (x ) = (x ) p (f g V sign((x )) pT g ^ pT g ^
Linear approaches

<0 If sign(p T g ) = sign(p T g ^ ) and > 0 suciently large, V The closed-loop system is robust against model errors
Outline
1
Linear approaches
3 4

Observers
State and output linearization

Introducing examples
Consider nonlinear the system 1 = 2x1 + ax2 + sin(x1 ) x 2 = x2 cos(x1 ) + u cos(2x1 ) x Take the new set of state variables z1 = x1 z2 = ax1 + sin(x1 ) The state equations become: 1 = 2z1 + z2 z 2 = 2z1 cos(z1 ) + cos(z1 ) sin(z1 ) + au cos(2z1 ) z The nonlinearities can then be canceled taking the new control v : u = 1 (v cos(z1 ) sin(z1 ) + 2z1 cos(z1 )) a cos(2z1 )
Linear approaches
x1 = z1 (z1 ) x2 = z2 sin a


The system then becomes linear: z = 2 1 0 z+ v 0 0 1
Taking v = 2z2 places the poles of the closed-loop in {2, 2} Hence, limt z = 0 Looking back to the transformation: x1 = z1 z2 sin(z1 ) x2 = a one also have limt x = 0 In the original coordinates, the control writes: u= 1 [2ax2 2 sin(x2 )cos(x1 ) sin(x1 )+2x1 cos(x1 )] a cos(2x1 )
Linear approaches


Consider nonlinear the system 1 = sin(x2 ) + (x2 + 1)x3 x

5 2 = x1 x + x3 2 3 = x1 x +u
Linear approaches
Take x1 as output: y = x1 When not imposed, the choice of the appropriate output is often delicate Compute the rst time derivative of y : =x 1 = sin(x2 ) + (x2 + 1)x3 y Compute the second time derivative of y :
5 2 = (x2 + 1)u + (x1 y + x3 )(x3 + cos(x2 )) + (x2 + 1)x1 m(x )

Taking u =
v m(x ) x2 +1
gives the linear system:
=v y

Potential problem if x2 = 1 that Here again, take the linear feedback v = y + 2y places the poles of the closed loop system in {1, 1} asymptotically converges to the origin Hence, y and y The state of the system is of dimension 3. Only two variables have been brought to the origin, what about the third one ? Write the system with the new coordinates , x3 ): (z1 , z2 , z3 ) = (y , y Linearized sub-system: Internal dynamics: 1 = z2 z 2 = v z
2 + u (z ) 3 = z1 z
Linear approaches


One has to check that the internal dynamics is stable. For this, we = 0 (which happens asymptotically): assume that y = y
0 0 0 5 z1 0 2 +z3 )(z3 + cos x2 ) + (x2 + 1)z1 ) x2 + 1
3 = z
2 z1 +
v ((
Linear approaches
z3 (z3 + cos x2 ) x2 + 1
3 z3 (z3 + 1) z3 : the internal If x2 and z3 are small enough, z dynamics is locally asymptotically stable The approach can be applied if and only if the internal dynamics is stable If the internal dynamics is unstable, the system is called non-minimum phase
Change the input Use additional inputs to stabilize the internal dynamics Other approach in the literature like approximate linearization


Introduction
Consider again nonlinear systems ane in the control: = f (x ) + g (x )u x y = h (x ) The linearization approach tries to nd
a feedback control law u = (x ) + (x )v and a change of variable z = T (x )
Linear approaches

that transforms the nonlinear systems into:
Input-State linearization
= Az + Bv z No systematic approach
Input-Output linearization
y (r ) = v Systematic approach

The SISO case
Take
= f (x ) + g (x )u x y = h (x )
The time derivative of y is then given by: h (f (x ) + g (x )u ) = Lf h(x ) + Lg h(x )u x If Lg h(x ) = 0, the control is = y u= 1 (Lf h(x ) + v ) Lg h(x )
Linear approaches

=v yielding y Otherwise (that is Lg h(x ) 0), dierentiate once more: = y Lf h (f (x ) + g (x )u ) = L2 f h(x ) + Lg Lf h(x )u x

The SISO case
If Lg Lf h(x ) = 0, the control is u= 1 (L2 f h (x ) + v ) Lg Lf h(x )

=v yielding to y ...
Linear approaches
Denition: Relative degree

There exists an integer r n such that Lg Lif h(x ) 0 for all 1 i {1, . . . , r 2} and Lg Lr h(x ) = 0 that is called the relative degree of f the system The control law u= 1
1 Lg Lr h (x ) f

(Lr f h (x ) + v )
yields the linear system y (r ) = v


The MIMO case
Take
= f (x ) + g1 (x )u1 + g2 (x )u2 x y1 = h1 (x ) y2 = h2 (x )
The relative degree {r1 , r2 } is then given by: Lg1 Lif1 h1 (x ) = Lg2 Lif1 h1 (x ) = 0 Lg1 Lif1 h2 (x ) = Lg2 Lif1 h2 (x ) = 0 i < r1 i < r2
Linear approaches

Assume that the Input-Output decoupling condition is satised, that is: rank
1 1 1 1 Lg1 Lr h1 (x ) Lg2 Lr h1 (x ) f f r2 1 r2 1 Lg1 Lf h2 (x ) Lg2 Lf h2 (x )
=2

The MIMO case
One has: 1 = Lf h1 (x ) + Lg1 h1 (x ) u1 + Lg2 h1 (x ) u2 y

=0 =0
1 = L2 y f h1 (x ) + Lg1 Lf h1 (x ) u1 + Lg2 Lf h1 (x ) u2
=0 =0
. . . y1 1
(r ) r1 1 1 1 1 = Lr h1 (x ) u1 + Lg2 Lr h1 (x ) u2 f f h1 (x ) + Lg1 Lf either =0 or =0
Linear approaches
Make the same for y2 : 2 = Lf h2 (x ) + Lg1 h2 (x ) u1 + Lg2 h2 (x ) u2 y

=0 =0

2 = L2 y f h2 (x ) + Lg1 Lf h2 (x ) u1 + Lg2 Lf h2 (x ) u2
=0 =0
. . . y2 2
(r ) r2 1 2 2 1 = Lr h2 (x ) u1 + Lg2 Lr h2 (x ) u2 f h2 (x ) + Lg1 Lf f either =0 or =0
Nonlinear Control
Master PSPI 2009-2010 112 / 174

The MIMO case
Hence, one has: y1 1 (r ) y2 2

(r )
r1 1 1 1 1 h1 (x ) Lg2 Lr h1 (x ) Lr f f h1 (x ) + Lg1 Lf r2 r2 1 r2 1 Lf h2 (x ) Lg1 Lf h2 (x ) Lg2 Lf h2 (x ) b (x ) A(x )
u1 u2
u
Thanks to the decoupling condition, one can take: u = A(x )1 b (x ) + A(x )1 v which gives two chains of integrators: y1 1 = v1
(r )
Linear approaches

y2 2 = v2
(r )
This approach can be extended to any output and input sizes

A block diagram representation
Linearization techniques can be interpreted as:
Linear approaches
Linear control Linearizing feedback
Nonlinear system
Linear loop


The general case
The internal dynamics or zero-dynamics is then of dimension n r ; its stability has to be checked For linear systems r =number of poles-number of zeros This approach took a rapid development in the 80s Maybe the only general approach for nonlinear system with predictive control Can be extended to MIMO systems with a decoupling condition Many extensions in particular with the notion of atness Power tool for path generation Non robust approach since it is based on coordinate changes that can be sti The control may take too large values Kills nonlinearities even if they are good

Linear approaches

Outline
1
Linear approaches
3 4

Observers
Backstepping
Main results
Theorem: backstepping design

= f (, v ) with f (0, 0) = 0 can asymptotically be stabilized with Assume that the system v = k (). Let V denote a C 1 Lyapunov function (denite, positive and radially V unbounded) such that f (, k ()) < 0 for all = 0. Then, the system = f (, ) = h(, ) + u with h(0, 0) = 0 is also asymptotically stabilizable with the control: u (, ) = h(, ) + with
1
Linear approaches
k V f (, ) + k () G (, k ())

G (, ) =
0
f (, k () + )d
The Lyapunov function W (, ) = V () + 1 k () 2

2
Observers
is then strictly decreasing for any (, ) = (0, 0)

Backstepping
Main results
The theorem can recursively be applied to

1 2 = = = . . . = = f (, 1 ) a1 (, 1 ) + b1 (, 1 )2 a2 (, 1 , 2 ) + b2 (, 1 , 2 )3
Linear approaches

n 1 n
an1 (, 1 , 2 , . . . , n2 ) + bn1 (, 1 , 2 , . . . , n2 )n1 an (, 1 , 2 , . . . , n1 ) + bn (, 1 , 2 , . . . , n1 )u
Forwarding
Strict-feedforward systems
Consider the class of strict-feedforward systems: 1 = x2 + f1 (x2 , x3 , . . . , xn , u ) x 2 = x3 + f2 (x3 , . . . , xn , u ) x . . . n1 = xn + fn1 (xn , u ) x n = u x Strict-feedforward systems are in general no feedback linearizable Backstepping is not applicable
Linear approaches

Forwarding
Forwarding procedure
n = un . At step 0: Begin with stabilizing the system x 2 Take e.g. un = xn and the corresponding Lyapunov function Vn = 1 2 xn At step 1: Augment the control law un1 (xn1 , xn ) = un (xn ) + vn1 (xn1 , xn ) such that un1 stabilizes the cascade n1 = xn + fn1 (xn , un1 ) x n = un1 x At step k: Augment the control law unk (xnk , . . . , xn ) = unk +1 (xnk +1 , . . . , xn ) + vnk (xnk , . . . , xn ) such that unk stabilizes the cascade nk = xnk +1 + fnk (xnk +1 , . . . , xn , unk ) . . .
Linear approaches

n1 = xn + fn1 (xn , unk ) x n = unk x
Forwarding
Forwarding procedure
The control law unk always exists for all k {1, . . . , n 1} 1 2 Vk = Vk +1 + xk + 2
The Lyapunov function at step k can be given by: xk ()fk (xk +1 (), . . . , xn ())d
0
Linear approaches

To avoid computations of the integrals, low-gain control can be used. It gave rise to an important litterature on bounded feedback control (Teel (91), Sussmann et al. (94)) Various extensions for more general system exist but are still based on the same recursive construction Forwarding can be interpreted with passivity, ISS or optimal control
Backstepping/Forwarding
An interpretation of the names
Linear approaches

Outline
1
Linear approaches
3 4

Observers
The X4 helicopter
X4 position stabilization
The aim is now to apply nonlinear control techniques to the stabilization problem of an X4 helicopter at a point Input-Ouput Linearization techniques can not be applied since the system is non minimum phase The backstepping approach is inspired from
S. Bouabdallah and R. Siegwart, Backstepping and sliding-mode techniques applied to an indoor micro quadrotor, at the EPFL (Lausanne, Switzerland) presented at the International Conference on Robotics and Automation 2005 (ICRA05, Barcelonna, Spain)
Linear approaches
The PVTOL control strategy comes from

A. Hably, F. Kendoul, N. Marchand and P. Castillo, Positive Systems, chapter: "Further results on global stabilization of the PVTOL aircraft", pp 303-310, Springer Verlag , 2006

The saturated control law is taken in:

N. Marchand and A. Hably, "Nonlinear stabilization of multiple integrators with bounded controls", Automatica, vol. 2147-2152, 2005. 41, no. 12, pp
Revealing example of what often happens in nonlinear: a solution comes from the combination of dierent methods
The X4 helicopter
A two steps approach
Linear approaches
i s p mv R J
= = v = = =
2 km Jr R s i
kgearbox Jr
|si | si +
mg e3 + R R J

Attitude control problem 0 r (s2 , s4 ) 0 + p (s1 , s3 ) i Ir y ( s 1 , s 2 , s 3 , s 4 ) i si
0 0 i F i ( si )
Position control problem
km Jr R
satU i (Ui )
The X4 helicopter: attitude stabilization

A backstepping approach
Linear approaches

The aim of this rst step is to be able to bring (, , ) to any desired conguration (d , d , d ) Attitude equations (with an appropriate choice of Euler angles): 1 = x2 x x 2 = a1 x4 x6 + b1 r 3 = x4 x 4 = a2 x2 x6 + b2 p x = x6 x 5 6 = a3 x2 x4 + b3 y x , , , ), a1 = J2 J3 , a2 = J3 J1 , with (x1 , . . . , x6 ) = (, , J1 J2 J2 1 1 1 a3 = J1J , b = , b = and b = 1 2 3 J1 J2 J3 3 The system would be trivial to control if (x2 , x4 , x6 ) was the control instead of (r , p , y ) This is precisely the philosophy of backstepping

Recall of backstepping main result

= f (, v ) with f (0, 0) = 0 can asymptotically be stabilized with Assume that the system v = k (). Let V denote a C 1 Lyapunov function (denite, positive and radially V unbounded) such that f (, k ()) < 0 for all = 0. Then, the system = f (, ) = h(, ) + u with h(0, 0) = 0 is also asymptotically stabilizable with the control: u (, ) = h(, ) + with
1
Linear approaches
k V f (, ) + k () G (, k ())

G (, ) =
0
f (, k () + )d
The Lyapunov function W (, ) = V () + 1 k () 2

2
Observers
is then strictly decreasing for any (, ) = (0, 0)


Dene rst the tracking error: z1 = x1 x1d z2 = x3 x3d z3 = x5 x5d It gives the subsystem 1 z 2 z 3 z
= d = d = d
Linear approaches

1d = x2 x 3d = x4 x 5d = x6 x
where, at this step, (x2 , x4 , x6 ) is the control vector

2 Take rst V1 = 1 2 z1 .

Along the trajectories of the system, the Lyapunov function gives: 1 = z1 (x2 x 1d ) V 1d 1 z1 with Hence taking as ctive control x2 = x 1 > 0 insures the decrease of V1 : 1 = 1 z 2 V 1 x1 will asymptotically converge to x1d ... unfortunately x2 is not the control and we have to build thanks to backstepping approach the r that will force x2 to make x1 converge to x1d For this, dene the tracking error for x2 : 1d 1 z1 ) z2 = x2 (x
input of subsystem in z1 we would like to put
Linear approaches

Master PSPI 2009-2010 129 / 174

One has 2 = a1 x4 x6 + b1 r x 1d + 1 x2 1 x 1d z
2 x 1 1 z
Recall we have to build the r that will force x2 to make x1 converge to x1d . For this, we use the formula of the backstepping theorem:

. . . u (, ) = h(, ) + V k f (, ) + k () G (, k ()) . . .
T
Linear approaches

with in our case: = z1 = z2 1d + 1 x2 1 x 1d h = a1 x4 x6 x

Nonlinear Control
1d 1 z1 k = x 1d f = x2 x 2 V = 1 z 2 1
Master PSPI 2009-2010 130 / 174

Or more simply, we construct it: W (z1 , z2 ) = V (z1 ) + 1 1d 1 z1 ) x2 ( x 2

2

1 2 2 = (z1 + z2 ) 2
Along the trajectories of the system: = z1 (x2 x 1d ) + z2 a1 x4 x6 + b1 r x 1d + 1 x2 1 x 1d W v b1 r gives: 1d + 1 x2 1 x 1d = a1 x4 x6 + b1 r x 1d 1 x2 + 1 x 1d = v a1 x4 x6 + x = z1 (x2 x 1d ) + z2 v W

control
Linear approaches
Taking v as new control variable:


Trying to write 1d x2 x = 1d + 1 z1 ) 1 z1 +? (x2 x

ideal case
dierence between real and ideal cases
1d + 1 z1 = 1 z1 + x2 x
z2
Linear approaches
Hence, it gives:
= 1 z 2 + z1 z2 +z2 v W 1
can be compensated with v

Taking: v = z1 2 z2 where 2 > 0 in order to insure = 1 z 2 2 z 2 < 0 W 1 2 Repeating this for and gives the wanted control law

Attitude stabilization
The control law given by 1d 1d 1 x2 + 1 x b1 r = z1 2 z2 a1 x4 x6 + x 3d 3d 3 x4 + 3 x b2 p = z3 4 z4 a2 x2 x6 + x 5d 5 x6 + 5 x 5d b3 y = z5 6 z6 a3 x2 x4 + x 1d + 1 z1 , z3 = x3 x3d , with z1 = x1 x1d , z2 = x2 x 3d + 3 z3 , z5 = x5 x5d and z6 = x6 x 5d + 5 z5 z4 = x4 x asymptotically stabilizes (x1 , x3 , x5 ) to their desired position (x1d , x3d , x5d ) This kind of approach appeared in the literature in the last 90s Now better approach based on forwarding are appearing taking into account practical saturation of the control torques
Linear approaches


Simulations: apply successive steps of 45 in roll, pitch and yaw

60 40 20 0 20 0 10 20 30 40 50 60 70
Linear approaches
60 40 20 0 20 60 40 20 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70


60 40 20 0 20 60 40 20 0 20 60 40 20 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0.5 1 0 1 2 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 0 10 20 30 40 50 60 70 1 0.5 0 10 20 30 40 50 60 70 0 1
Adjusting the s: with i = 0.2
Linear approaches

Roll, pitch and yaw answers
r , p and y controls
Too many oscillations

60 40 20 0 20 60 40 20 0 20 60 40 20 0 2 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 2 2 0 10 20 30 40 50 60 70 0 0 10 20 30 40 50 60 70 5 2 0 10 20 30 40 50 60 70 0 5
Adjusting the s: with i = 2
Linear approaches

Seems to be a good choice

60 100 0 100 50 0 50 50 0 50
Adjusting the s: with i = 10

40 20 0 20 60 40 20 0 20 60 40 20 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Linear approaches

The controls are too large, may be two fast to consider the rotors as instantaneous

Linear approaches

Successive steps of 45 in roll, pitch and yaw with i = 2


The saturation of the possible control torques r , p , y may be problematic On the X4 system, the maximum speed of the rotors is given by the solution of:
2 km smax + kgearbox Rsmax2 km U = 0
Linear approaches

which gives: smax = 604 rad.s1 The maximum roll and pitch torques are when one rotor is at rest, the other one at is maximum speed (we assume that the rotation is in only one direction):
2 rmax ,p = lbsmax = 0.31 N.m
The maximum yaw torque is obtained when two opposite rotors turn at the smax the two others are at rest:
max 2 y = 2smax = 21.2 N.m

Adapt the parameters to the saturation:

Take larger 5,6 than the 1,...,4 Take small enough i to fulll the saturation constraint
Linear approaches


60 40 20 0 20 60 40 20 0 20 60 40 20 0 50 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0.5 50 0 10 20 30 40 50 60 70 0 0 10 20 30 40 50 60 70 1 0.5 0 10 20 30 40 50 60 70 0 1
Adjusting the s: with 1,...,4 = 0.4 and 5,6 = 8
Linear approaches

The controls are still too large

60 40 20 0 20 60 40 20 0 20 60 40 20 0 50 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0.5 50 0 10 20 30 40 50 60 70 0 0 10 20 30 40 50 60 70 1 0.5 0 10 20 30 40 50 60 70 0 1
Adjusting the s: with 1,...,4 = 0.2 and 5,6 = 8
Linear approaches

Oscillations are now present

Nonlinear Control N. Marchand References Outline
60 0.5 0 0.5 0.5 0 0.5 50 0 50
Taking 1,...,4 = 0.2 and 5,6 = 8 and applying it on the system with saturations, it gives:
40 20 0 20 60 40 20 0 20 60 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Introduction
Linear approaches

40 20 0 20 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Seems to work

Linear approaches

The control law with the saturated system

The X4 helicopter: position control

Simplication of the problem
First use the attitude control to adjust the X4 in the right direction:
Linear approaches


There exists a relation between (, , ) and (, , ) (a rotation about the yaw axis) Use the attitude control to drive and keep and to the origin Take now the system in the plane:
Linear approaches


This problem is referred in the literature as the PVTOL aircraft stabilization problem (Planar Vertical Take O and Landing aircraft) The generic equations are: = sin()u x = cos()u 1 y = v
u y v
Linear approaches

where -1 represent the normalized gravity, v represents the equivalent control torque.

A saturation based control
Dene z (z1 , z2 , z3 , z4 , z5 , z6 )T The system becomes:
)T , y, y , , (x , x
Linear approaches
Translational part: 1 = z2 z Rotational part: 2 = u sin(z5 ) 5 = z6 z z 3 = z4 6 = v z z 4 = u cos(z5 ) 1 z d such that: The idea is to use v to drive z5 to z5 z5d with <
1 2
arctan(

r1 ) r2 + 1
(tuning parameter) and () = max(min(, 1), 1): r1 = (z2 ) 2 (z1 + z2 ) r2 = (z4 ) 2 (z3 + z4 )

When z5 reaches z5d and applying the thrust control input u u=

2 + (r + 1)2 r1 2
Linear approaches
the translational subsystem takes the form of two independent second order chain of integrators z1 = z2 z3 = z4 z2 = r1 z4 = r2 If we could prove that r1 = (z2 ) 2 (z1 + z2 ) r2 = (z4 ) 2 (z3 + z4 ) insures (z1 , z2 , z3 , z4 ) 0 Then, it will follow that once (z1 , z2 , z3 , z4 ) = 0, z5d = 0 and 5d = z6d = 0 hence also (z5 , z6 ) 0 z


Remain two problems: Prove that r1 = (z2 ) 2 (z1 + z2 ) r2 = (z4 ) 2 (z3 + z4 ) brings (z1 , z2 , z3 , z4 ) to zero 5d = z6d ) Build a control so that (z5 , z6 ) tends to (z5d , z
Linear approaches


A saturated control law for linear systems
An integrator chain is dened by:

0 . . . . = . x . . . . 1 .. . 0 1 .. . ...
=:A
0 ...
... .. . .. . .. . ...
0 0 . . . . . . u . x + . 0 . 0 1 1 0
=:B
(11)
Linear approaches
Problem: stabilizing (11) with uu u is a positive constant and u


First compute
n i =1
( + i ) = p0 + p1 + + pn1 n1 + n T T Tx with
Apply the coordinate change y :=

n i i =1
n n1
Linear approaches
Tn2 T1
= B = (A + pn1 I )B = (A2 + pn1 A + pn2 I )B . . . = (An1 + pn1 An2 + + p1 I )B

The system becomes normalized in

0 n1 n2 . . . 0 0 n2 . . . . .. .. . . . . 0 . . . ... 0 0 ... ... 0 . . y . 0
1 . . . . . v . . . . 1
with
n i i =1
n i i =1
Master PSPI 2009-2010 152 / 174
Nonlinear Control

Theorem: An ecient saturated control

Let (n) denote the only root of n 2 + 1 = 0 in ]0, 1[. Then for all with 0 < < (n), u= u
n
Linear approaches

n i i =1 i =1
ni +1 sat1 (yi )
globally asymptotically stabilizes the integrator chain.

Sketch of the proof:

2 Assume that |yn | > 1 and take Vn := 1 2 yn Then
n = yn sat1 (yn ) yn 2 sat1 (yn1 ) + + n sat1 (y1 ) V

= ||<2 ++n
Linear approaches
So V is decreasing if > 2 + + n 1 2 + n > 0 < ( n ) yn joins [1, 1] in nite time and remains there During that time, yn1 to y1 can not blow up Repeating this scheme for yn1 to y1 gives y B (1) after some time In B (1), the system is linear and stable GAS

The construction of this feedback law is based on feedforward

Remain two problems: Prove that r1 = (z2 ) 2 (z1 + z2 ) r2 = (z4 ) 2 (z3 + z4 ) brings (z1 , z2 , z3 , z4 ) to zero direct application of the saturated control law 5d = z6d ) Build a control so that (z5 , z6 ) tends to (z5d , z take: 5d ) (z6 z 5d ) 2 ((z5 z5d ) + (z6 z 5d )) v = (z that work applying the saturated control law on the 5d variables z5 z5d and z6 z
Linear approaches

The X4 helicopter
Second step: position control
Linear approaches

(, , , x , y , z ) = ( pi , pi , pi , 5, 5, 5) 2 2 4 , , x ,y ,z ) = (0, 0, 0, 1, 2, 0) (,
The X4 helicopter
Second step: position control
Linear approaches

Zoom on the 10 rst seconds with a snapshot every second

Initial conditions: (, , , x , y , z ) = ( pi , pi , pi , 5, 5, 5) 2 2 4 , , x ,y ,z ) = (0, 0, 0, 1, 2, 0) (,
Outline
1
Linear approaches
3 4

Observers
Nonlinear observers
yd
k (x )
Controller
x ^
Observer
1
Linear approaches
Modelization
To get a mathematical representation of the system Dierent kind of model are useful. Often:
a simple model to build the control law a sharp model to check the control law and the observer

Design the state reconstruction: in order to reconstruct the variables needed for control Design the control and test it Close the loop

Linear systems: Simple observer (t ) = Ax (t ) + Bu (t ) x y (t ) = Cx (t )

Observability given by rank(C , CA, . . . , CAn1 ) Once the observability has been checked, dene the observer: (t ) = Ax x ^ ^ + Bu (t ) + L(^ y (t ) y (t )) y ^ (t ) = C x ^ (t ) with L such that (eig(A + LC )) < 0. Then x ^ tends to x asymptotically with a speed related to eig(A + LC )
Linear approaches


Linear systems (cntd.):

convergence of the observer Dene the error e (t ) := x ^(t ) x (t ). Then: (t ) = Ae (t ) + L(^ e y (t ) y (t )) = (A + KC )e (t ) Hence, if (eig(A + LC )) < 0, limt x ^ (t ) = x (t ) Separation principle: the controller and the observer can be designed separately, if each are stable, their association will be stable
Linear approaches

Linear systems (cntd.): Kalman lter

(t ) = Ax x ^ ^(t ) + Bu (t ) + L(^ y (t ) y (t )) y ^ (t ) = C x ^ (t )
The way to chose L proposed by Kalman is: AP + PAT PC T W 1 CP + V + P = 0 W = WT > 0 L = PC T W 1 with > 2 A or V = V T > 0. enables to tunes the speed of convergence of the observer The observer can also be computed as L = S 1 C T W 1 where AT S + SA C T W 1 C + SVS + S = 0
Linear approaches

Time-varying linear systems:

(t ) = A(t )x (t ) + B (t )u (t ) x y (t ) = C (t )x (t ) Kalman lter: = A(t )^ x ^ x (t ) + B (t )u (t ) + L(t )(^ y (t ) y (t )) y ^ (t ) = C (t )^ x (t ) with: = AP + PAT PC T W 1 CP + V + P P

T P (0) = P0 = P0 >0
Linear approaches

W = WT > 0 L = PC T W 1 > 2 A or V = V T > 0 enables to tunes the speed of convergence of the observer
The observer can also be computed as L = S 1 C T W 1 where

= AT S + SA C T W 1 C + SVS + S S
T S (0) = S0 = S0 >0
The observer is optimal in the sense that it minimizes

t 0
Linear approaches
[C ()z () y ()]T W 1 [C ()z () y ()] + [B ()u ()]T V 1 [B ()u ()] d +

1 (x0 x ^0 )T P0 (x0 x ^0 )

The observer is also optimal for

x is aected by a white noise wx of variance V y is aected by a white noise wy of variance W wx and wy are uncorrelated
Nonlinear systems:
(t ) = f (x (t ), u (t )) x y (t ) = h(x (t )) Observability
Linked with the notion that for two trajectories of the system x1 and x2 dened on [0, t ], we must have
t
Linear approaches
h(x1 ()) h(x2 ()) d > 0 if x1 = x2

0

Linked with the input: = x 0 0 u x 0 y= 1 0 x
is observable for any u (t ) = 0 but not for u (t ) = 0
Extended Kalman lter (EKF):

= f (x (t ), u (t )) L(t )(^ x ^ y (t ) y (t )) y ^ (t ) = h(^ x (t )) with = AP + PAT PC T W 1 CP + V + P P

T P (0) = P0 = P0 >0
Linear approaches
W = WT > 0 L = PC T W 1 > 2 A or V = V T > 0 h f (^ x (t ), u (t )) C = (^ x (t )) A= x x No guarantee of convergence (except for specic structure conditions)

Luenberger-Like observers
For a system of the form:
(t ) = Ax (t ) + (Cx (t ), u ) x y (t ) = Cx (t ) with (A, C ) observable, if A KC is stable, an observer is:

(t ) = Ax x ^ ^(t ) + (y (t ), u (t )) K (C x ^(t ) y (t )) For a system of the form: (t ) = A0 x (t ) + (x (t ), u (t )) x y (t ) = C0 x (t ) with (A0 , C0 ) are in canonical form, if A0 K0 C0 is stable, suciently large, global lipschitz and i (x , u ) = 0 for j i + 1 xj an observer is: (t ) = A0 x x ^ ^(t ) + (^ x (t ), u (t )) diag(, 2 , . . . , n )K0 (C0 x ^(t ) y (t ))
Linear approaches

Kalman-Like observers
For a system of the form: (t ) = A(u (t ))x (t ) + B (u (t )) x y (t ) = Cx (t ) an observer is: (t ) = A(u (t ))^ x ^ x (t ) + B (u (t )) K (t )(C x ^(t ) y (t )) with = A(u (t ))P + PAT (u (t )) PC T W 1 CP + V + P P
T P (0) = P0 = P0 > 0, W = W T > 0
Linear approaches

> 2 A(u (t )) L = PC T W 1
or V = V T > 0
Kalman-Like observers (cntd.)

For a system of the form:
(t ) = A0 (u (t ), y (t ))x (t ) + (x (t ), u (t )) x y (t ) = C0 x (t ) with: 0 0 1 a12 (u , y ) .. . 0 bounded an1n (u , y ) 0
A0 (u , y )
Linear approaches
C0 i (x , u ) xj
= =
0 for j i + 1

an observer is: = A(u , y )^ x ^ x + (^ x , u ) diag(, 2 , . . . , n )K0 (t )(C0 x ^ y) with = A(u (t ))P + PAT (u (t )) PC T W 1 CP + P P
T P (0) = P0 = P0 > 0, W = W T > 0
> 2 A(u (t )) L = PC T W 1
and large enough

Not mentionned: Optimal observers (robust, very ecient and easy to tune but costly) Based on something like:
t1
Linear approaches
x ^ = Arg min
x () t0
h(x ()) y ()) d

Sliding mode observers
Observers
Application to the attitude estimation
Application: attitude estimation

9 sensors:
3 triax accelerometers 3 triax gyrometers 3 triax magnetometers
The accelerometers give: bacc = C (q )( a

acceleration
+ g ) + acc
gravity noise
Linear approaches

2 where C (q ) = (q0 q T q )I3 + 2(q q T q0 q ) is called the Rodrigues matrix, that is the rotation from the xed frame to the mobile one. The magnetometers give:
bmag = C (q )hmag + mag

noise
where hmag are the coordinates of the magnetic eld in the xed frame.
Observers
The gyrometers give: bgyr = + gyr1 + gyr1

noise bias
Linear approaches

where hmag are the coordinates of the magnetic eld in the xed frame. The bias drift is the main error and it deteriorates the accuracy of the rate gyros on the low frequency band. We take: bgyr = + gyr1 + gyr1 1 = gyr1 + gyr2 gyr1
noise
Observers
Nonlinear observer for attitude estimation

An observer of the attitude can be: = 1 (bgyr q ^ ^ gyr1 + K1 )^ q 2 = T 1 ^ ^ K2 where T is a diagonal matrix of time constant, Ki are positive denite matrices, is given by: = qe sign(qe0 ) where qe = (qe0 , qe ) is the quaternion error between the estimate q ^ and a direct projection obtained with bmag and bacc
Linear approaches

Application to the estimation = qeattitude sign(qe ) o

1 qe = q Observers (t) qms (t) = [qe0 qe ]T
(28) (29)
Block diagram of the observer
It is possible to demonstrate that the error model for (25) and T T (26) has the equilibrium state [qe b ] = [1 0 0 0 0 0 0]T and is exponentially stable [19].
Fig.
Eule chan rate obse
Linear approaches

Fig. 3.
Block diagram showing the attitude estimation
The complet schema of the attitude observer is given in Fig. 3. Remark 2: suppose that the rigid body Master has a slow angular Nonlinear Control PSPI 2009-2010 174 / 174

Nonlinear To Linear System Project

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Nonlinear To Linear System Project

Загружено:

Авторское право:

Доступные форматы

Control of Nonlinear Systems

Nonlinear Control N. Marchand

CNRS Associate researcher

Nicolas.Marchand@grenoble-inp.fr Control Systems Department, gipsa-lab

Stability Nonlinear approaches

Master PSPI 2009-2010

Observers N. Marchand (gipsa-lab)

Probably the best book to start with nonlinear control

Nonlinear systems S. Sastry - Springer Verlag, 1999

Mathematical Control Theory - E.D. Sontag - Springer, 1998

Constructive nonlinear control - Sepulchre et al. - Springer, 1997 More

Nonlinear control systems - A. Isidori - Springer Verlag, 1995

Applied Nonlinear control - J.J. Slotine and W. Li - Prentice-Hall, 1991

Stability Nonlinear approaches

Observers N. Marchand (gipsa-lab) Nonlinear Control Master PSPI 2009-2010 2 / 174

Stability Nonlinear approaches

Stability Nonlinear approaches

Stability Nonlinear approaches

Observers N. Marchand (gipsa-lab) Nonlinear Control Master PSPI 2009-2010 5 / 174

Stability Nonlinear approaches

Some preliminary vocable

We consider a dynamical system:

Stability Nonlinear approaches

Observers N. Marchand (gipsa-lab) Nonlinear Control Master PSPI 2009-2010 7 / 174

The 4 steps to control a system

Stability Nonlinear approaches

Linear dynamical systems

Stability Nonlinear approaches

The model can be obtained

Observers N. Marchand (gipsa-lab)

both give h(s ) or (A, B , C , D ) and hence the model.

Linear dynamical systems

= Ax + Bu x y = Cx (+Du ) Nice properties:

Stability Nonlinear approaches

Observers N. Marchand (gipsa-lab)

Master PSPI 2009-2010 10 / 174

Why nonlinear control ?

All physical systems are nonlinear because of

Stability Nonlinear approaches

Observers N. Marchand (gipsa-lab)

Nonlinear dynamical systems

In this course, only:

PDE Partial dierential equation (trac, ow, etc.): 0 0 = =

Stability Nonlinear approaches

The model can be obtained

Observers N. Marchand (gipsa-lab) Nonlinear Control Master PSPI 2009-2010 12 / 174

Nonlinear dynamical systems

Stability Nonlinear approaches

Nonlinear dynamical systems

References Outline Introduction

Stability Nonlinear approaches

Null stabilization problem

Observers N. Marchand (gipsa-lab)

Some strange behaviors of nonlinear systems

Stability Nonlinear approaches

No proportionality between the input and the output

Observers N. Marchand (gipsa-lab) Nonlinear Control Master PSPI 2009-2010 15 / 174

Some strange behaviors of nonlinear systems

Van der Pol oscillator 1 = x2 x 2 = x1 h(x1 )x2 x

Oscillations: tunes the limit cycle

= 1, x (0) = (5, 5) = 0.1, x (0) = (5, 5)

= 1, x (0) = (5, 5) = 0.1, x (0) = (5, 5)

Stability Nonlinear approaches