Feedback Linearisation: A Linearising Approach for 1

Affine Class of Control Systems

O. T. C Nyandoro
School of Electrical and Information Engineering
University of the Witwatersrand, Johannesburg
o.nyandoro@ee.wits.ac.za

Abstract
These notes present f eedback linearisation (IOF BL) as a tool for linearising a class of nonlinear systems. In
particular IOF BL is presented as a technique for linearising af f ine systems. Examples are provided in formulating
IOF BL and two methods are provided to implement IOF BL. While both methods are similarly based on integration
of the output function, the second method utilises Lie Algebra to formulate IOF BL. Lastly P ole placement is
used as a method to design a tracking controller for the IOF BL linearised system. References used for these notes
are mainly [1] and [2], [3], [4], [5], [6]

I. METHOD 1: I NPUT-O UTPUT F EEDBACK L INEARISATION (IOFBL)

I NTRODUCTION - D IFFERENTIATION BASED FBL F ORMULATION
Consider the nonlinear SISO system which is af f ine i.e. linear in the control, with n states:
ẋ = f(x) + g(x)u (1)
y = h(x) (2)
with the objective being to make the output y(t) track a desired (pre-determined) trajectory yd (t) while
keeping all the n states x bounded. The challenge in meeting such a design goal is that the output y is only
indirectly related to the input u. Hence designing a control − law is not easy for such a nonlinear such.
Input-output feedback linearisation IOF BL refered to as F BL in these notes seeks to establish a direct
relationship between the output y and the control input u. This idea constitutes the basis for F BL.

A. Differentiation Based F BL
EXAMPLE 1 Consider the system 3rd − order SISO system

ẋ1 = sin(x2 ) + x3
ẋ2 = x51 + x3 (3)
ẋ3 = x21 + u
y = x1 (4)
Differentiating (4) to obtain

ẏ = ẋ1 = sin(x2 ) + x3 (5)

gives ẏ which is still not directly related to u (i.e u does not appear in the equation for ẏ). So we
differentiate (5) yet again to obtain ÿ:
d(ẏ) d(ẋ1 )
ÿ = = = cos(x2 )ẋ2 + ẋ3 = cos(x2 )(x51 + x3 ) + x21 + u (6)
dt dt
i.e.

ÿ = cos(x2 )(x51 + x3 ) + x21 + u (7)

The input u is now directly related to the output y, i.e the control input u appears in a derivative of the
output y. We thus formulate a new control input unew equal to the right-handside of (7):
 2

unew = cos(x2 )(x51 + x3 ) + x21 + u (8)

to give a new linear state equation

ÿ = unew
(9)
for which pole placement can now be used to obtain the control law for the input unew .
The system control input is then derived from (8) as

u = unew − cos(x2 )(x51 + x3 ) − x21 (10)

.
So F BL can be summarised as the process:
• differentiate the output equation and check if the derivative of y is directly/explicitly related to the
input u
• if the above is not satisfied then continue differentiating the output y successively until a particular
derivative of y is explicitly related to the input.
• once a derivative of y is directly related to u then formulate a new linear input as the right hand side
of the resptive derivative of y that is directly related to u
• linear control methods such as pole placement can then be used to formulate feedback gains for unew
from which u can then be obtained.
REMARK The number of differentiations of y, r are called the relative degree, and usually r ≤ n.

EXAMPLE 2
ẋ1 = −x1 + e2x2 u
ẋ2 = 2x1 x2 + sinx2 + 0.5u (11)
ẋ3 = 2x2
y = h(x) = x3 (12)

ẏ = 2x2 (13)
ÿ = 2ẋ2 = 2(2x1 x2 + sinx2 ) + 0.5u = unew (14)
F BL gives a linearised system of relative degree r = 2 for the system of order n = 3

II. P OLE P LACEMENT FOR S TABLE L INEARISED S YSTEM

For asymptotically stable closed loop poles for the linearised system of relative degree r and with
linearised input unew , let

ẋnew = Axnew + Bunew use (15)

unew = Kxnew obtain (16)
0 = |sI − A + BK| = (s − µ1 )(s − µ1 )...(s − µr ) (17)
Or phrased differently but still with the same result, inear pole place design for unew :
unew = −kr−1 y (r−1) − . . . − k1 ẏ − k0 y (18)
gives the equation in the s − plane

sr + kr−1 sr−1 + . . . + k1 s + k0 = 0 (19)

Choice of ki in (19) such that all poles are strictly in the left-half plane gives a locally asymptotically
stable closed loop linearised system with input un ew and the linearised states being the series of r output
derivatives. Finally u is obtained from the equation for unew from the rth derivative of y.
 3
III. P OLE P LACEMENT FOR L INEARISED T RACKING S YSTEM
Following on from the above stabilised pole placement the tracking system is similarly designed. For
asymptotically stable tracking of a desired trajectory yd (t) for the linearised system of relative degree r and
with linearised input unew , let
(r−1) T
αd = [yd ẏd . . . yd ] (20)
and the tracking error vector is:

α̃(t) = α − αd (21)
where the output vector α:

α = [y ẏ . . . y (r−1) ]T (22)
Hence the linearised output is similar to the stabilised pole placement result albeit for error dynamics to
go to zero:
(r)
unew = yd − kr−1 α̃(r−1) − . . . − k1 α̃˙ − k0 α̃ (23)
gives the equation in the s − plane

FN

Damper
ω
r
xs Sprung Mass τ

Active
Actuator Spring v Bω
xu Unsprung Mass
Kt
xr Fz
FY
road FX
d

Fig. 1. Quarter Car Braking Model

1 35

0.9

30

0.8

25
0.7
Braking velocities (m s-1)

0.6
µ

20
and

0.5
λ

15
0.4

0.3
10

0.2

0.1

0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
time (s) time (s)

Fig. 2. λ and µ (left) and on (right) Velocities for ABS with Suspension ρ Tracking
 4

sr + kr−1 sr−1 + . . . + k1 s + k0 = 0 (24)

making the error dynamics vector go to zero
• usually r < n meaning that less state dynamics are available in the linearised system than in the
original system.
• if r < n then some state dynamics are hidden in the state model by the linearising F BL process
• if r < n then care should be taken in applying F BL to ensure that the hidden state is bounded, (this
MUST be specially noted) and checked for
• the hidden states are obtained and zero dynamic analysis is then performed to check the boundedness
of the hidden states

IV. M ODELLING ABS WITH S USPENSION S YSTEM E FFECTS

For detailed modelling of ABS (without suspension effects) see [2].

ẋ1 −µ(λ)FN /M − µ(λ)kw x4 /M − Cx x21 /M

=
ẋ2 µ(λ)FN r/Iw + µ(λ)kw x4 r/Iw − Bω/Iw − ub /Iw
=
ẋ3 =
x5
ẋ4 =
x6
ẋ5 −ks (x3 − x4 )/ms − ksnl (x3 − x4 )3 /ms − bs (ẋ3 − ẋ3 )/ms
=
−bsym
s |(ẋ3 − ẋ3 )| /ms + us /ms (25)
ẋ6 = ks (x3 − x4 )/mu + ksnl (x3 − x4 )3 /mu + bs (ẋ3 − ẋ4 )/mu
+bsym
s |(ẋ3 − ẋ4 )| /mu − kw x4 /mu − us /mu
y = [y1 y2 ]T = [λ γ]T (26)
where the ABS states are x = [x1 x2 x3 x4 x5 x6 ]T , x1 = v, x2 = ω, x3 = xs , x4 = xu , x5 = ẋs ,
x6 = ẋu . y is the output column vector consisting of the output slip (27) and output suspension travel (28).
ub and us are the braking torque τb as braking control input and active suspension force as the suspension
control input respectively.

v − rω
yabs = λ = (27)
v

ysus = γ = xs − xu (28)
The ABS model can thus be defined in the state-space form as the non-linear multi-input multiple output
system:
ẋ = f (x) + gu(t) (29)
y = h (x) = [λ γ]T (30)
where
g = [0 1/Iw 0 0 0 0]
[0 0 0 0 1/ms − 1/mw ]T (31)
u = [ub us ]T (32)
Highly non-linear f(x) consists of all terms of (25) except for the terms in us and ub which then appear
in g(x). The actuator dynamics and similar suspension system dynamics are left out of the state equations
mainly to simplify the state equations.

Multi control of both suspension and ABS can be done separately by formulating tracking control for
the suspension then tracking control for the ABS. Sample test results for suspension ρ tracking and ABS λ
tracking are shown in the sample graphs Fig.2.
 5
V. METHOD 2: L IE N OTATION BASED FBL F ORMULATION
Let h(x) be a scalar function of the state x and f(x) be a vector field in <n . Then differential geometry
defines the gradient of h(x) by ∇h(x):
∂h
∇h = (33)
∂x
(33) represents a row vector of elements (∇h)j = ∂h/∂xj |j = 1 . . . n
similarly the Jacobian ("gradient" for a vector field) of the vector field f(x) denoted as ∇f:
∂f
∇f = (34)
∂x
Yet again (34) represents a n-row vector of n-column vector elements (i.e nxn) matrix of elements
(∇h)ij = ∂fi /∂xj |i = 1 . . . n, j = 1 . . . n.

Definition: Lie Derivative [1]

The Lie derivative of h with respect to f is a scalar function defined by Lf h = ∇hf
From differential geometry the directional derivative of h in the direction of f is represented by the Lie
derivative.
Repeated Lie derivatives can recursively apply Lie derivatives:

L0f h = h (35)
Lif h = Lf (Li−1 i−1
f h) = ∇(Lf h) f or i = 1, 2 . . . (36)
If g is another vector field in <n using Lie notation Lg Lf h(x) is a scalar function:

Lg Lf h = ∇(Lf h)g (37)

Formulation of F BL is thus possible using Lie notation. From (1)

ẋ = f(x) + g(x)u (38)

y = h(x) (39)
taking derivatives

∂h ∂h
ẏ = ẋ = (f(x) + g(x)u) = Lf h (40)
∂x ∂x
∂[Lf h]
ÿ = ẋ = L2f h (41)
∂x
(iii) ∂[L2f h]
y = ẋ = L3f h (42)
∂x
...
y (r) = Lrf h(x) + Lg Lr−1 f h(x)u = unew (43)
More formally from (38)
∂h
ẏ = ẋ = ∇h(f + gu) = Lf h(x) + Lg h(x)u = unew (44)
∂x
and if the input appears in (44) then unew holds in which case as formulated earlier
1
u= (−Lf h + unew ) (45)
Lg h
If the input does not appear in (44) then repeated derivatives are taken and specifically because the input
does not appear in (44) then all coefficients of u must necessarily be zero i.e. (Lg h = 0) so this term does
not need to be considered in subsequent differentiations. Consequently differentiating ẏ after taking out
terms in u results in
 6

ÿ = L2f h(x) + Lg Lf h(x)u = unew (46)

If again the input does not appear then Lg Lf h(x) = 0 and we differentiate (46) until after r differentiations
the input does appear and the system has a relative degree in which case unew holds true and generally
1
u= (−Lrf h + unew ) (47)
Lg Lr−1
f h
obtain from the r-output differentiations giving the linear relation:
y (r) = Lrf h(x) + Lg Lr−1
f h(x)u = unew (48)
As noted earlier generally r ≤ n.
Finally from previous section stabilisation problems result in
1
u= (−Lrf y − kr−1 y (r−1) − . . . − k1 ẏ − k0 y) (49)
Lg Lr−1
f y
while tracking problems result in
1
u= (−Lrf α̃ − kr−1 α̃(r−1) − . . . − k1 α̃˙ − k0 α̃) (50)
Lg Lr−1
f α̃
where the ki s in (50) and (49) are chosen as specified earlier in the pole placement sections

A. FBL Formalisation

∂h
µ̇1 = ẏ = ẋ = Lf h
∂x
∂[Lf h]
µ̇2 = ÿ = ẋ = L2f h
∂x
∂[L2f h]
µ̇3 = y (iii) = ẋ = L3f h
∂x
... = ...
∂[Lr−2 h]
µ̇r−1 = y (r−1) = f
ẋ = Lr−1
f h (51)
∂x
µ̇r = y (r) = Lrf h(x) + Lg Lr−1 f h(x)u = unew (52)
Note that generally the relative degree, r: 1 ≤ r ≤ n. For cases where r = n then IOF BL is a total
state transformation from system state xn to new state µn i.e. xn → µn .
However if r < n then IOF BL transforms µr and the other n − r states namely ψn−r are hidden. Hence
IOF BL still transforms xn → zn where the new state z = [µr ψn−r ]T . While µr is known via Method 1
of differentiation or method via Lie derivative approach the states ψn−r are not part of the controller design
and are regarded as hidden. Their behaviour may cause internal instability hence their stability needs to be
found or deduced.

B. Hidden States
Usually it suffices to find a candidate solution for ψn−r from
Lg ψj = 0; 1≤j ≤n−r (53)
∂z
Additionally the Jacobian matrix, ∂x must not be singular as verified by that the inverse of this Jacobian
matrix exists.
Lastly the stability of the hidden states can be analysed from the zero-dynamics analyses. This can be
done by first setting all r states µr = 0 and then noting the stability behaviour of ψ̇ and/or its effect on the
output when the input u is set to zero.
 7
C. Examples
•
   2x 
−x1 e 2
ẋ = 2x1 x2 + sinx2  +  1/2  u (54)
2x2 0
y = h(x) = x3 (55)
• Consider ABS with IOF BL

R EFERENCES
[1] J. J. Slotine and W. Li, "Applied Non Linear Control," Prentice Hall, New York, 1991.
[2] J.O. Pedro and O.T.C. Nyandoro and S. John, "Neural Network Based Feedback Linearisation Slip Control of an Anti-Lock Braking
System," Proc. of the 7th Asian Control Conference, Sep 2009, pp 1251-ï¿ 21 1257, Hong Kong, China, ISBN:978-89-956056-9-1.
[3] O.T.C. Nyandoro and J.O. Pedro and B. Dwolatzky, "Control-Scheduling Codesign: Real-time Optimal Braking of Wheeled Vehicles",
7th South African Conference on Computational and Applied Mechanics - SACAM10, Jan 2010, Pretoria, South Africa.
[4] O.T.C. Nyandoro and J.O. Pedro and B. Dwolatzky and O. Dahunsi, "Control-scheduling codesign of real-time optimal braking of antilock
braking systems", Second African Conference on Computational Mechanics - An International Conference - AfriCOMP11, Jan 2011, Cape
Town, South Africa.
[5] O.T.C. Nyandoro and J.O. Pedro and O. Dahunsi and B. Dwolatzky, "Linear Slip Control Formulation for Vehicular Anti-Lock Braking
System with Suspension Effects", Proc. of the 19th IFAC World Congress, Aug 2011, Milan, Italy.
[6] O.T.C. Nyandoro and J.O. Pedro and B. Dwolatzky and O. Dahunsi, "State Feedback Based Linear Slip Control Formulation for Vehicular
Antilock Braking System", Proc. of the World Congress in Engineering 2011 - International Conference of Applied and Engineering
Mathematics (Selected for Best Paper), Jul 2011, London, United Kingdom.