Nonlinear Control Systems 1. - Introduction To Nonlinear Systems

Nonlinear Control Systems
1. - Introduction to Nonlinear Systems
Dept. of Electrical Engineering
Department of Electrical Engineering

University of Notre Dame, USA
EE60580-01
Dept. of Electrical Engineering (ND) Nonlinear Control Systems 1. - Introduction to Nonlinear SystemsEE60580-01 1 / 54
Course Overview
The course has three modules

Advanced Tools for ODEs
Stability Concepts
Nonlinear Controller Synthesis
Text - lecture notes that are transcribed from a variety of texts and
monographs
Homework (once a week) 1-3 problems - We’ll arrange recitation
sessions for HW groups (2-3 students). HW due the day after
recitation. Recitations are optional after the first week.
Two exams - Midterm/Final - closed book.
Course Overview
Novel features of this course

1 Emphasis of relationships between tools and concepts, rather than
detailed discussion of proofs (proofs are in notes).
2 Topics on bifurcation and structural stability
3 Use more modern approach of treating Lyapunov functions as stability
certificates. Examines computational methods for certifying system
stability.
4 Emphasis on three stabilization methods; constructive methods based
on CLF, older geometric methods use in feedback linearization, and
passivity-based methods.
5 Introduces relatively recent ideas regarding Model-free control
schemes that provide a ”hybrid systems” approach to the regulation
of nonlinear processes.
Dynamical Systems
Rather than starting with an ODE as the system model, let us

consider a more general formulation in which the process is seen as a
mathematical system.
Σ = (X , G , U, φ) where X is a topological space (state space), G is a
strongly ordered group of time instants, U is a linear space of input
signals defined on G , and φ : G × X × U → X is a transition map.
We assume the transition map is continuous and
φ(0, x, u) = x (identity)
φ(s + t, x) = φ(t, φ(s, x, u[0,s] ), u[s,s+t] )
We call x : G → X a state trajectory or orbit. The value x(t) denotes

the system ”state” at time instant t ∈ G .
Dynamical Systems
If there is no input (i.e. u ≡ 0), then this is an unforced system
(X , G , φ) and with the given assumptions we know that the orbits do
not intersect and partition the phase space into equivalence classes.
This justifies the notion of ”state”.
Principle of Superposition
φ(t; x, αu1 + βu2 ) = αφ(t; x, u1 ) + βφ(t; x, u2 )
means the system is linear. You studied this in linear systems theory
(EE550) when G with either Z (discrete-time) or R
(continuous-time).
We focus on systems that do not satisfy the principle of
superposition. These systems exhibit a wide range of interesting
behaviors that do not appear in linear systems.
The objective of the first couple lectures is to present examples
introducing these interesting behaviors. These behaviors will provide
the motivation for our subsequent study of nonlinear control systems.
Discrete-time System based on Logistic Map
Let us consider a discrete-time system Σ = (X , G , φ) where

X = [0, 1], G = Z, and φ : [0, 1] → [0, 1] is given by
φ(x) = ax(1 − x) (1)
with a being a positive real constant. This is called the logistic map.
Orbits are generated by the recursive equation
xk+1 = axk (1 − xk ), x0
where a > 0 and {xk }∞

k=0 is the orbit.
Models population of organism in a box of fixed size. When
population is low, then there is abundant food/space and xk is
increasing. When population is high, there is a shortage of
food/space and xk is decreasing.
Logistic Map
cobweb plot population state trajectory
1 1
0.9 0.9
0.8
a=2 0.8
0.7 0.7
0.6 0.6
population
f(x(k))
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 4 6 8 10 12 14 16 18 20
x(k) generation (k)
Figure: Cobweb Graph and State Trajectory for Logistic System with a = 2 and
x0 = 0.2
Logistic Map
One can computationally examine what happens if the food supply, a,

is increased without expanding the logistic map’s positive support (i.e.
the living space). This gives rise to a number of scenarios
Extinction: (0 ≤ a < 1) - In this case the graph of Φ(x) never crosses

the 45 degree line formed by y = x. x[k] decreases with each
generation and asymptotically approaches 0. In other words, for this
range of a the population goes extinct due to a lack of food.
Stable Population: (1 ≤ a < 3) - In this case the 45 degree line for
y = x intersects the graph y = Φ(x) at two points xu∗ ≡ 0 and
xs∗ ≡ 1 − 1a . One can readily verify that the point xs∗ = 1 − 1a is an
asymptotically stable equilibrium in the sense that x[k] → xs∗ as
k → ∞. This case, therefore, corresponds to a scenario in which the
population achieves a “stable” fixed level.
Logistic Map
√
Limit Cycling Population: (3 < a ≤ 1 + 6 = 3.449) - In this case
the population goes through a “boom and bust” cycle or what will be
later referred to as a limit cycle. Fig. 2 shows the cobweb plot and
state trajectory for the case when a = 3.4. In this figure the
population bounces between two different sizes on each generation.
1 1
0.9 0.9
0.8
h=3.4 0.8
0.7 0.7
0.6 0.6
population
f(x(k))
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 4 6 8 10 12 14 16 18 20
x(k) generation (k)
Figure: Cobweb Graph and State Trajectory for Logistic System with h = 3.4
and x0 = 0.2
Logistic Map
Road to Chaos:√ (3.449 < a ≤ 3.570) - For a just greater than
3.449 = 1 + 6, the period 2 limit cycle begins doubling until
a > 3.570 at which point the population’s state trajectory becomes
chaotic as shown in Fig. 3. The term ”chaos” formally means that
the future states vary in a discontinuous manner with the initial
condition x0 .
1 1
0.9 0.9
0.8
h=3.9 0.8
0.7 0.7
0.6 0.6
population
f(x(k))
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 100 200 300 400 500 600 700 800 900 1000
x(k) generation (k)
Figure: Cobweb Graph and State Trajectory for Logistic System with h = 3.9
and x0 = 0.2
Dept. of Electrical Engineering (ND) Nonlinear Control Systems 1. - Introduction to Nonlinear Systems
EE60580-01 10 / 54
Nonlinear Continuous-time Oscillator
A nonlinear dynamical system, S = (X , G , U, φ), is said to be

continuous-time if the index set G is the set of reals, R.
Continuous-time systems can also exhibit a range of behaviors that
include convergence to fixed points, limit cycles, and chaos.
The Duffing oscillator satisfies a second order differential equation of
the form
ẍ + γ ẋ − ω 2 x + x 3 = Γ cos(Ωt) (2)
where ω, γ, , and Γ are all positive real constants.
EE60580-01 11 / 54
Duffing Oscillator
The Duffing equation (2) may

be viewed as a forced oscillator
whose spring provides a
restoring force that is a cubic
function of the system state.
When < 0 then the spring is
a ”hardening” spring and when
> 0, then this spring is
”soft”.
EE60580-01 12 / 54
Duffing Oscillator
Numerically integrating equation (2) forward in time shows a periodic limit

cycle when γ = 0.1, = 0.25, ω = 1, Γ = 0.5 and Ω = 2.
time series phase space Poincaré section (GAM=0.5)
0.5 2
0.4
0.4 1.5
0.3 0.3
1
0.2 0.2
0.5
0.1 0.1
0 0 0
-0.1 -0.1
-0.5
-0.2 -0.2
-1
-0.3 -0.3
-0.4 -1.5
-0.4
-0.5 -2
80 100 120 140 160 180 1.8 1.9 2 2.1 2.2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
Figure: State trajectory (left), phase plane trajectory (middle), and Poincaré
section (right) for forced Duffing oscillator (2) with γ = 0.1, = 0.25, ω = 1,
Γ = 0.5, and Ω = 2.
EE60580-01 13 / 54
Poincare Section
Poincaré section provides a

convenient way of viewing the
behavior of periodic state
trajectories. The method fixes
a cross section, Σ, of the phase
space and determines a map F
that maps a state x ∈ Σ onto
the state trajectory’s first
return, F (x), to the cross
section.
EE60580-01 14 / 54
Duffing Oscillator Chaos
When we change Γ = 1.5 in the Duffing oscillator, the Poincaré section
shows a much more complex structure. This structure is sometimes called
a chaotic or strange attractor for the system.
time series phase space Poincaré section (GAM=1.5)
3 3
2.5
2 2 2
1.5
1 1
1
0.5
0 0
0
-1 -0.5
-1
-1
-2 -1.5
-2
-2
-3
80 100 120 140 160 180 -3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2
Figure: State trajectory (left), phase plane trajectory (middle), and Poincaré
section (right) for forced Duffing oscillator (2) with γ = 0.1, = 0.25, ω = 1,
Γ = 1.5, and Ω = 2.
EE60580-01 15 / 54
Nonlinear Regulation Problems
Consider an input-output system whose state trajectory, x : R → Rn ,

system output y : R → Rm , and control input u : R → Rp satisfy the
following set of equations,
ẋ(t) = f (x, u)
y = h(x, u)
u = k(x)
We assume that the functions f : Rn × Rp → Rn and

h : Rn × → Rp → Rm are known.
Given a desired output signal, yd : R → Rm , the objective is to find
the state feedback controller k : Rn → Rp such that the output signal
y (t) → yd (t) as t → ∞.
EE60580-01 16 / 54
Regulation through Local Linearization
Consider a system ẋ = f (x, u) where u is fixed and let φ denote the

system’s flow.
A point x ∗ ∈ Rn is called an equilibrium point if f (x ∗ , u) = 0 for the
given u ∈ Rm . This equilibrium point for the unforced system (i.e.
u = 0) is said to be hyperbolic if the eigenvalues of the Jacobian
matrix , ∂x∂f
(x ∗ , 0) , evaluated at x ∗ has no eigenvalues with zero real
parts.
The Hartman-Großman theorem establishes that the flow φ in a
suitably small neighborhood of a hyperbolic equilibrium is
topologically equivalent to the flows of the linearized system that
satisfies

∂f ∗ ∗ ∂f ∗
ẋ = (x , 0) (x − x ) + (x , 0) u (3)
∂x ∂u
= A(x − x ∗ ) + Bu
EE60580-01 17 / 54
Topological equivalence means there is a smooth invertible

state-space transformation between the flows of the nonlinear system
and its linearization that preserves the direction of time.
This suggests that if one were to design a state feedback controller
matrix, K , such that the control signal u = K (x − x ∗ ) asymptotically
stabilizes the linearized system about this equilibrium point, then we
should achieve adequate regulation of the nonlinear system.
EE60580-01 18 / 54
Nonholonomic Control of Two-wheeled Robot
Let us see how well this ”linearized” control works for a two-wheeled
robotic vehicle
The “plant” is a two-wheeled robotic vehicle. Let F denote the force
applied along the body’s x-axis and T be the torque about the
vehicle’s center of mass.
The control vector is u(t) = [F (t), T (t)]T . The state variables are
the plant’s center of mass, x and y , the body angle, θ, the velocity,
vx , and the angular rate, ω.
EE60580-01 19 / 54
The first step in developing such a state feedback control is to find

the linearization in equation (3) for our two-wheeled cart.
We start by introducing the new tracking variables z1 = x − xd ,
z2 = y − yd , z3 = θ − θd , z4 = vx − vd , and z5 = ω where
(xd (t), yd (t)) is the trajectory we want our vehicle to track in the
plane, θd (t) is the direction of the desired trajectory’s velocity vector,
and vd (t) is the magnitude of that desired velocity vector.
With this change of variables our system equations in Fig. ?? become
(z4 + vd ) cos(z3 + θd ) − ẋd

     
ż1 0 0
 ż2   (z4 + vd ) sin(z3 + θd ) − ẏd   0 0 
      u1
 ż3  = 
   z5 − θ̇d + 0 0
 

 u2
 ż4   −v̇d   1 0 
ż5 0 0 1
= F (x) + G (x)u
EE60580-01 20 / 54
Computing the Jacobian matrix for F and Gu yields the following

linearized system equation
    
ż1 0 0 −vd sin(θd ) cos(θd ) 0 z1
 ż2   0 0 vd cos(θd ) sin(θd ) 0   z2 
    
 ż3  =  0 0 0 0 1   z3 
 
  
 ż4   0 0 0 0 0   z4 
ż5 0 0 0 0 0 z5
 
0 0
 0 0 
  u1
+ 0 0 
 
 1 0  u2
0 1
= Az + Bu
EE60580-01 21 / 54
We can then use a number of methods that design stabilizing
controllers for this system. In particular, we’ll compute the linear
quadratic regulator (LQR) that find the state gains, K , such that the
controller u = Kz minimizes the cost functional
Z ∞
J[u] = (z T z + u T u)dτ
0
This controller was simulated in the following MATLAB script with
the desired trajectory

2πt
ẋd (t) = 50 sin , xd (0) = 0
50

4πt
ẏd (t) = 50 cos , yd (0) = 0
50
The LQR control was recomputed at each time instant using the
desired reference trajectory states. We indeed obtain tracking of the
desired reference trajectory, though the vehicle’s initial transient
shows some significant oscillation while it is picking up speed.
EE60580-01 22 / 54
250
desired traj
vehicle traj
200
initial position
final position
150
direction of time
100 vehicle tracking
50 direction of time
y
-50
-100
-150
-200
-250
-100 0 100 200 300 400 500 600 700 800
x
Figure: (left) MATLAB script - (right) trajectories for linearized control with
(x0 , y0 ) = (50, 0)
EE60580-01 23 / 54
A limitation of the preceding

linearization approach is that 200
desired traj
vehicle traj
the topological equivalence is 150
only local (i.e. in a 100
neighborhood of the 50
direction of
time
equilibrium point). This 0
y
suggests that if we were to -50
vehicle is lost
start the vehicle further away -100
from the desired reference -150
trajectory then our control -200

-400 -200 0 200 400 600 800
strategy might fail. This indeed x
is the case for our system.
EE60580-01 24 / 54
Regulation through Global Linearization
One way of overcoming this limitation is to adopt a more

sophisticated feedback linearization method that uses feedback to
force the system to appear as a linear dynamical system.
This feedback linearization is a state transformation that transforms
the original system states onto a state vector consisting of the desired
tracking outputs and their derivatives.
The advantage of this approach is that if that state transformation is
“global”, then the controls we develop for this “feedback” linearized
system are also global and can thereby ensure asymptotic tracking of
the reference trajectory for any initial vehicle condition.
EE60580-01 25 / 54
In the feedback linearization approach we will find it convenient to
introduce a change of control variables in which
Z t
u1 (t) = F (t) = v1 (s)ds
0
u2 (t) = T (t) = v2 (t)
The original control, u1 = F , is then treated as another system state,
thereby extending the state vector of the original system.
With this change of control variable we obtain the following state
equations for our cart,
     
x vx cos θ 0 0
 y   vx sin θ   0 0 
     
d  θ =
  ω   0 0  v1
+ 
 vx   F
dt      0 0  v2
  
 ω   0   0 1 
F 0 1 0
EE60580-01 26 / 54
We now introduce a state transformation which is obtained by taking
the derivatives of the tracking error.
z1 = x − xd z2 = ẋ − ẋd , z3 = ẍ − ẍd
z4 = y − yd , z5 = ẏ − ẏd , z6 = ÿ − ÿd
The differential equations for these components are then readily
computed as
ż1 = vx cos θ − ẋd = z2
ż2 = F cos θ − vx ω sin θ − ẍd = z3
...
ż3 = v1 cos θ − vx v2 sin θ − (2F ω sin θ + vx ω 2 cos θ) − x d
ż4 = vx sin θ − ẏd = z5
ż5 = F sin θ + vx ω cos θ − ÿd = z6
...
ż6 = v1 sin θ + vx v2 cos θ + (2F ω cos θ − vx ω 2 sin θ) − y d
EE60580-01 27 / 54
These equations have the form of two chains of integrators driven by

the inputs into states z3 and z6 . This means we can rewrite the above
differential equations in the following form,
ż1 0 1 0 0 0 0 z1 0 0
      
ż2 0 0 1 0 0 0 z2 0 0
       ...
ż3 0 0 0 0 0 0 z3 1 0 −(2F ω sin θ + vx ω 2 cos θ) − x d
      
= + ...
      
ż4 0 0 0 0 1 0 z4 0 0 2F ω cos θ − vx ω 2 sin θ − y d
    
      
 ż5   0 0 0 0 0 1  z5   0 0 
ż6 0 0 0 0 0 0 z6 0 1

cos θ −vx sin θ v1
+
sin θ vx cos θ v2
ż = Az + E (α + ρv )
where A is the linear matrix representing the chain of integrators, α

and ρ are matrices whose components are functions of the original
system states, F , ω, θ, and vx .
EE60580-01 28 / 54
Regulation Through Global Linearization
Note that if we select the control v to have the form
...
−1 xd
v =ρ −α + ... y + Kz (4)
d
Then the resulting state equation is given by

ż = (A + EK )z (5)
where

T 0 0 1 0 0 0
E =
0 0 0 0 0 1

k11 k12 k13 k14 k15 k16
K =
k21 k22 k23 k24 k25 k26
The important thing to note here is that equation (5) is a linear
differential equation and so if we can select K so that A + EK is a
Hurwitz matrix, then we would have globally stabilized our vehicular
system using the control in equation (4).
EE60580-01 29 / 54
300
desired traj
x0=(50,0)
250
x0=(250,250)
direction of
initial position
200
time final position
150 direction of time

perfect tracking
100
50
y
direction of
0
time
-50
-100
-150
-200
0 100 200 300 400 500 600 700 800
x
Figure: Feedback Linearized Controller (left) script - (right) trajectories
EE60580-01 30 / 54
Bursting in Adaptive Systems
Adaptive control systems are systems that adapt the controller’s gains
in response to how well the controlled system is actually behaving.
Even when this approach is used to adapt a linear dynamical system,
the overall system will be highly nonlinear since the control signal is a
product of the gain and the system state; both of which in turn are
time-varying.
This nonlinearity makes can make it difficult to ensure the adaption
rule is well behaved (i.e. stable) and it sometimes leads to unexpected
behaviors such as bursting; a phenomenon that occurs when an
adaptive system suddenly exhibits a burst of oscillatory behavior
before returning back to normal.
Bursting is a qualitative behavior found in biological systems and
turbulent fluid flow. The objective of this section is to illustrate
bursting in a simple adaptive control problem and show how the
nonlinearities in that system give rise to the bursting phenomena.
EE60580-01 31 / 54
Bursting in adaptive linear systems
As a simple example, let us consider a linear plant whose state
x : R → R satisfies
ẋ = ax + bu (6)
in which a and b are unknown system constants but in which we
assume we can measure the system state x(t).
The objective is to select a control signal u : R → R so that x tracks
a reference model whose state xm : R → R satisfies
ẋm = −am xm + bm r (7)
where am > 0, bm > 0, r (t), and xm (t) are all known.
The control input, u, is generated by the following equation
u(t) = k1 x(t) + k2 r (t) (8)
where k1 and k2 are “control gains”.
The objective is to select these gains so that the model tracking error,
e = x − xm , asymptotically goes to zero as time goes to infinity.
EE60580-01 32 / 54
Let us first note that the tracking error satisfies the differential
equation
ė = ẋ − ẋm
= ax + b(k1 x + k2 r ) + am x − bm r
Note that if we select k1∗ = − a+a

b
m
and k2∗ = bm
b , then
ė = ax + b(k1∗ x + k2∗ r ) + am x − bm r

a + am bm
= ax + b − x+ r + am x − b m r (9)
b b
= −am (x − xm ) = −am e
Since am > 0, this would imply that for k1 = k1∗ and k2 = k2∗ that
e(t) asymptotically goes to 0 as t → ∞.
EE60580-01 33 / 54
This particular choice for the control gains, however, requires that one
know what a and b are ahead of time. Since these system parameters
are unknown, our actual control law uses a k1 and k2 that are not
equal to the ”optimal” values.
For this non-optimal choice, the modeling error, e, satisfies
ė = ax + b(k1 x + k2 r ) + am x − bm r
= ax + b((k1 − k1∗ )x + (k2 − k2∗ )r ) + bk1∗ x + bk2∗ r + am − bm r
We can then use equation (9) to reduce the above equation to
ė = −am e + b(k1 − k1∗ )x + b(k2 − k2∗ )r (10)
which more clearly shows how deviations in k1 and k2 from their

optimal values impact the model tracking error.
EE60580-01 34 / 54
Bursting in Adaptive Linear Systems
In model reference adaptive control (MRAC), one seeks an adaptation

rule that adjusts the controller parameters k1 and k2 to ensure model
tracking.
This adaptation rule may be expressed as a differential equation
k̇1 = φ1 (k1 , x, r , e), k̇2 = φ2 (k2 , x, r , e) (11)
where φ1 and φ2 are the functions we need to determine.

The adaptation rule in equation (11) takes as input the current plant
output, x, the reference input r and tracking error e to adjust the
gains k1 and k2 .
EE60580-01 35 / 54
To see what a plausible form for these adaptation functions should
be, let us consider a cost function
1 b b
V (e, k1 , k2 ) = e 2 + (k1 − k1∗ )2 + (k2 − k2∗ )2
2 2γ 2γ
where γ > 0 is a constant.
Note that V ≥ 0 and let us consider the time derivative of this cost,
V̇ , as e, k1 , and k2 follow the error dynamics in equation (10) under
the adaptation rule (11).
In particular, if one can show that V̇ ≤ 0, then since V is bounded
below by zero, then V (t) must be a monotone decreasing function of
time that converges to a limit point.
So let us compute V̇
∂V ∂V ∂V
V̇ = ė + k̇1 + k̇2
∂e ∂k1 ∂k2
b b
= e ė + (k1 − k1∗ )k̇1 + (k2 − k2∗ )k̇2 (12)
γ γ
EE60580-01 36 / 54
Inserting equation (10) into (12) yields,
1 1
V̇ = −am e 2 + b(k1 − k1∗ )(xe + k̇1 ) + b(k2 − k2∗ )(xr + k̇2 )
γ γ
Note that if we select
k̇1 = φ1 (k1 , x, e, r ) = −γxe

(13)
k̇2 = φ2 (k2 , x, e, r ) = −γxr
then we can ensure V̇ < 0 for all e 6= 0 and this would suggest V (t)
is a decreasing function of time.
The adaptation rule in equation (13) is sometimes called the MIT rule
and it represents one of the earliest adaptive control laws proposed for
aerodynamic systems in the 1960’s.
EE60580-01 37 / 54
Various forms of the MIT rule have been used, but one well known
variation takes the form
x
k̇1 = −γ θ+x 2e
(14)
k̇2 = −γxr
This is sometimes called the normalized MIT rule since the size of the
k1 update is normalized by the size of the current system state.
Let us now look at simulations of the normalized MIT rule and see
what happens. In this simulation, we let the plant and reference
model be
ẋ = 1.8x + 2u
ẋm = −3x + 3r
EE60580-01 38 / 54
Let us first simulate this with γ = 1 and θ = 0.1 using the normalized MIT
rule from equation (14). The top plot shows e(t) → 0 as time goes to
infinity. The lower plot shows k1 and k2 and the optimal gains (k1∗ and k2∗ )
for this example.
3
x, plant
2 xm, model
-1
-2
-3
0 50 100 150 200 250 300 350 400 450 500
time
3
k1
2 k2
k1star
k2star
1
-1
-2
-3
0 50 100 150 200 250 300 350 400 450 500
time
Figure: Simulation of Normalized MIT rule for ẋ = 1.8x + 2u and ẋm = −3x + 3r
with no noise in the measured plant state.
EE60580-01 39 / 54
Bursting in Adaptive Linear System
An important thing to note about the simulation is that there is no
modeling error in the linear plant and there is no measurement noise.
To see how our proposed adaptive control works under these
conditions, let us assume that the plant state, x, satisfies the ODE,
ẋ = 1.8x + 0.1x 2 + 2u
y = x + 0.1ν
where ν is a zero mean white noise process with unit variance.
We now use the measured output y , rather than the true state in our
feedback controller and adaptation rules,
ê = y − xm
u = k1 y + k2 r
y
k̇1 = −γ ê
0.1 + y 2
k̇2 = −γ êr
EE60580-01 40 / 54
The simulation results show t there is a burst in which the system
becomes highly oscillatory and then recovers tracking again.
8
x, plant
6 xm, model
bursting
4
phenomenon
2
-2
-4
0 50 100 150 200 250 300 350 400 450 500
time
4
k1
2 k2
k1star
k2star
0
-2
-4
-6
-8
0 50 100 150 200 250 300 350 400 450 500
time
The lower plot shows that k1 and k2 have a linear drift and bursting occurs
when these gains change sign.
EE60580-01 41 / 54
Bursting in Nonlinear Systems
There are numerous other systems where the bursting phenomenon
plays an important functional role.
This occurs in the dynamics of cell membranes where the ”burst”
corresponds to an impulse representing the transmission of
information down a nerve cell’s axon.
A simplified version of the Fitzhugh-Nagumo model for nerve
conduction.
v3
v̇ = v− − w + (y + I )
3
ẇ = φ(v + a − bw )
The variable v represents the cell membrane’s potential, w represents
a ”gating” variable, and I represents the stimulating current
(assumed to be constant).
For a suitable constant I , the membrane potential exhibits a sustained
oscillation as shown in the top plot of Fig. 9.
EE60580-01 42 / 54
Bursting in Nonlinear Systems
Bursting occurs when a small modulation in the driving currrent I + y is
introduced. The burst is triggered when the current perturbation, y ,
exceeds a given threshold.
Voltage Potential with Constant Current, I
3
-1
-2
-3
-4
0 50 100 150 200 250 300 350 400 450 500
Voltage Potential with Slow variation in I

3
bursting
2
-1
-2
-3
-4
0 50 100 150 200 250 300 350 400 450 500
Figure: Membrane Potential v , (top) constant applied current I - (bottom)

modulated current I + y
EE60580-01 43 / 54
Reachability in Nonlinear Control Systems
Consider a linear time-invariant system
ẋ(t) = Ax(t) + Bu(t), x(0) = 0 (15)
We’re interested in finding a control input, u that transitions the
system state from x(0) = 0 to the state x(T ) = x1 where T > 0.
The state x1 is said to be reachable if there exists a finite time T > 0
and a control input u : [0, T ] → Rm such that
Z T
x1 = e A(T −τ ) Bu(τ )dτ
0
The system is said to be reachable if every state in Rn is reachable
from the origin.
The necessary and sufficient condition for the reachability of the
linear system in equation (15) is that
rank A AB A2 B · · · An−1 B = n

(16)
EE60580-01 44 / 54
In particular, one usually examines reachability with respect to

nonlinear systems that satisfy the following state equations
ẋ(t) = A(x) + B(x)u
where A(x) is an n × n matrix of functions, aij : Rn → Rn , in which

A(0) = 0 and B(x) is an n × m matrix of functions, bij : Rn → Rm .
Since we know the flows about a hyperbolic equilibrium of the
nonlinear system are topologically equivalent to flows about the
linearized system’s equilibirum, this suggests that one might be able
to use the reachability condition in equation (16) to determine
whether or not the nonlinear system is reachable.
This is, however, not as simple as it seems on the surface. A simple
example can be used to illustrate how nonlinearity complicates this
question.
EE60580-01 45 / 54
Let us consider the rigid body dynamics for a satellite controlled by
two gas jets.
It will be convenient to define an inertially fixed reference frame
defined a dextral set of 3 vectors, {e1 , e2 , e3 }, and a reference frame
attached to the satellite’s body defined by another dextral set of
vectors {r1 , r2 , r3 }.
We define the direction cosine matrix of the vehicle as a 3 by 3 matrix
R such that Rei = ri for i = 1, 2, 3. This matrix may be obtained
through the sequence of rotations {θ1 , θ2 , θ3 } about the body frame
{r1 , r2 , r3 } assuming that the body was originally aligned with the
inertial frame {e1 , e2 , e3 }. These angles are called Euler angles.
For i = 1, 2, 3 let ci denote cos(θi ) and si denote sin(θi ) then we
obtain the following explicit expression for the direction cosine matrix,
 
c2 c3 −c2 s3 s2
R =  s1 s2 c3 + s3 c1 −s1 s2 s3 + c3 c1 −s1 c2 
−c1 s2 c3 + s3 s1 c1 s2 s3 + c3 s1 c1 c2
Note that the orderNonlinear
of theControl
Dept. of Electrical Engineering (ND) rotations
Systems 1. can be changed.
- Introduction to Nonlinear Systems
EE60580-01 46 / 54
The body’s velocity vector, ω is defined about the 3 body axes,
{r1 , r2 , r3 }.
The direction cosines and the angular velocities satisfy the following
set of ordinary differential equations
m
X
J ω̇ = S(ω)Jω + bi ui (17)
i=1
Ṙ = S(ω)R
where J is the inertia matrix, {bi }m i=1 , are vectors characterizing how
the gas jets, ui , for i = 1, 2, 3 impact the body’s angular acceleration,
ω̇, and
 
0 ω3 −ω2
S(ω) =  −ω3 0 ω1 
ω2 −ω1 0
EE60580-01 47 / 54
We could also have defined these angles through a different sequence of

rotations. Fig. 10 shows how these rotations would have occurred if we
had used the sequence {θ3 , θ1 , θ2 } rather than the 1 − 2 − 3 sequence.
Figure: 3-1-2 Rotation of Euler Angles
EE60580-01 48 / 54
We are more interested in how the three Euler angles change over
time since these are tied to the inertial frames. So rather than
characterizing the attitude’s time rate of change through its direction
cosine matrix (i.e., Ṙ = S(ω)R), we determine the time derivatives
for the Euler angles θi , for i = 1, 2, 3.
For the 1 − 2 − 3 sequence of body rotations it can be shown that (?,
?)
 
c2 0 s2
θ̇ =  s2 s1 /c1 1 −c2 s1 /c1  ω (18)
−s2 /c1 0 c2 /c1
Equations (17,18) will therefore be used in our example.
EE60580-01 49 / 54
For this example, we will see whether or not the origin is reachable
from an initial state where θ1 = −0.3, θ2 = θ3 = 0, and
ω1 = ω2 = ω3 = 0 using only two gas jets aligned with the vehicle’s
pitch and yaw axes. We’ll assume that the inertia matrix is
J = diag(0.1, 1.0, 1.0).
So the vehicle is starting out at rest aligned with the reference frame
but with a roll angle of set of −0.3 radians. Since the two gas jets are
aligned to the pitch and yaw axis then in equation (17) we know
m = 2, b1 = [0, 1, 0]T and b2 = [0, 0, 1]T .
EE60580-01 50 / 54
Let us first look at the linearization of this system about the set point
ω = 0 and θ = 0. We let the linearized system’s state vector be
z1 = ω1 , z2 = ω2 , z3 = ω3 , z4 = θ1 , z5 = θ2 , and z6 = θ3 .
It can be readily shown that for this fixed point the linearized
system’s state equations are
      
ż1 0 0 0 0 0 0 z1 0 0

 ż2  
  0 0 0 0 0 0 
 z2  
  1 0 

 ż3  
= 0 0 0 0 0 0  z3  
+ 0 1  u1
 u2 = Az + Bu
  

 ż4  
  1 0 0 0 0 0 
 z4  
  0 0 
 ż5   0 1 0 0 0 0  z5   0 0 
ż6 0 0 1 0 0 0 z6 0 0
EE60580-01 51 / 54
If we use the reachability condition in equation (16) then we see that
rank A AB A2 · · · A5 B = 4 < 6

which means the linearized system is not reachable.

In particular, this means that we cannot find a control that moves the
roll angle, θ1 , to zero. For the linearization this should be obvious
since we see there is no control action that impacts ż1 .
The same, however, may not be said of the original nonlinear system.
EE60580-01 52 / 54
Reachability of Nonlinear Control Systems
Let us consider the following control sequence
p t−0.1 − p t−1.8 − p t−3.1 t−5.25

0.4 0.4 0.4 +p 0.4
u=
p t−0.1 t−1.8 t−6.1 t−7.725

0.4 − p 0.4 − p 0.4 + p 0.4
where p(t) = 1 for 0 < t < 1 and is zero elsewhere.
1
u1
0.5 Pitch Control Jet
0
-0.5
-1
0 5 10 15
0.5 Yaw Control Jet
-0.5
-1
0 5 10 15
Figure: Gas jet commands used to null the roll angle with only pitch and yaw
control
EE60580-01 53 / 54
When both the pitch and yaw rates accerate at the same time, there is an
induced roll rate which we will null out the initial roll angle. We can then
use the pitch and yaw jets to then return the other angles to zero. So even
though the origin is not reachable from the initial attitude using the
linearized system, it is reachable using the original nonlinear dynamics.
50
roll axis
-50
0 5 10 15
50
pitch axis
-50
0 5 10 15
50
yaw axis
-50
0 5 10 15
EE60580-01 54 / 54

Nonlinear Control Systems 1. - Introduction To Nonlinear Systems

Uploaded by

Copyright:

Available Formats

Nonlinear Control Systems 1. - Introduction To Nonlinear Systems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Nonlinear Control Systems 1. - Introduction To Nonlinear Systems

Uploaded by

Copyright:

Available Formats

Nonlinear Control Systems

1. - Introduction to Nonlinear Systems

Dept. of Electrical Engineering

Department of Electrical Engineering

The course has three modules

Novel features of this course

Rather than starting with an ODE as the system model, let us

We call x : G → X a state trajectory or orbit. The value x(t) denotes

Let us consider a discrete-time system Σ = (X , G , φ) where

φ(x) = ax(1 − x) (1)

where a > 0 and {xk }∞

One can computationally examine what happens if the food supply, a,

Extinction: (0 ≤ a < 1) - In this case the graph of Φ(x) never crosses

A nonlinear dynamical system, S = (X , G , U, φ), is said to be

where ω, γ, , and Γ are all positive real constants.

The Duffing equation (2) may

Numerically integrating equation (2) forward in time shows a periodic limit

Poincaré section provides a

Consider an input-output system whose state trajectory, x : R → Rn ,

We assume that the functions f : Rn × Rp → Rn and

Consider a system ẋ = f (x, u) where u is fixed and let φ denote the

Topological equivalence means there is a smooth invertible

The first step in developing such a state feedback control is to find

(z4 + vd ) cos(z3 + θd ) − ẋd

Computing the Jacobian matrix for F and Gu yields the following

A limitation of the preceding

only local (i.e. in a 100

equilibrium point). This 0

from the desired reference -150

trajectory then our control -200

is the case for our system.

One way of overcoming this limitation is to adopt a more

These equations have the form of two chains of integrators driven by

where A is the linear matrix representing the chain of integrators, α

Then the resulting state equation is given by

150 direction of time

Figure: Feedback Linearized Controller (left) script - (right) trajectories

Note that if we select k1∗ = − a+a

We can then use equation (9) to reduce the above equation to

ė = −am e + b(k1 − k1∗ )x + b(k2 − k2∗ )r (10)

which more clearly shows how deviations in k1 and k2 from their

In model reference adaptive control (MRAC), one seeks an adaptation

k̇1 = φ1 (k1 , x, r , e), k̇2 = φ2 (k2 , x, r , e) (11)

where φ1 and φ2 are the functions we need to determine.

Inserting equation (10) into (12) yields,

k̇1 = φ1 (k1 , x, e, r ) = −γxe

Voltage Potential with Slow variation in I

Figure: Membrane Potential v , (top) constant applied current I - (bottom)

In particular, one usually examines reachability with respect to

ẋ(t) = A(x) + B(x)u

where A(x) is an n × n matrix of functions, aij : Rn → Rn , in which

We could also have defined these angles through a different sequence of

Figure: 3-1-2 Rotation of Euler Angles

Equations (17,18) will therefore be used in our example.

If we use the reachability condition in equation (16) then we see that

which means the linearized system is not reachable.

0.5 Yaw Control Jet

You might also like

where ω, γ, , and Γ are all positive real constants.