You are on page 1of 21

Preprints submitted to IJC Aldeen & Sharma

ESTIMATION OF STATES, FAULTS AND UNKNOWN

DISTURBANCES IN NONLINEAR SYSTEMS

M. Aldeen and R. Sharma*

Department of Electrical and Electronic Engineering

The University of Melbourne

Australia.

* Corresponding author

ABSTRACT

In this paper, we extend existing theory on nonlinear unknown input observer

design to a wider class of nonlinear systems where the systems are subject to

unknown input and output disturbances and experience faults. The approach used is to

decouple the faults and unknown disturbances from the rest of the system through a

series of transformations on state and output equations. Once total disturbance

decoupling is achieved, an appropriate observer for the disturbance free part of the

nonlinear system is proposed and designed. The designed observer is, subsequently,

used for estimation of unknown inputs, outputs, and fault signals. Two examples are

given for illustration.

Keywords: Nonlinear Observer, Fault Detection, Unknown Inputs.

Page 1 of 21
Preprints submitted to IJC Aldeen & Sharma

1. INTRODUCTION

Observer based techniques have been the subject of extensive research over the

past four decades. For linear systems, the problem of observer design has reached

maturity and various techniques have been proposed, see for example (Guan and Saif

(1991), Hou and Muller (1992,1994), Keonig and Patton (1999)). One early technique

is due to (Kurek 1983) where necessary and sufficient conditions and their

constructive proofs for the existence of unknown input observer (UIO) for linear

multivariable systems are presented. (Park and Stein 1988) investigated the problem

of simultaneously estimating the state and unknown inputs using a reduced-order

observer. The observer design is based on transforming the system into a special form

obtained through linear transformations. In (Yang and Wilde 1988) a full order UIO

design ocedure based on matrix operations is presented for linear systems where the

unknown input appears in the state only. The method takes into consideration the

basic system constraints, such as ranks of disturbance and output matrices.

For nonlinear systems, however, there still remain many open issues. This is

particularly the case when such systems experience unknown disturbances in both the

input and output, which are a source of erroneous information. Most of real life

industrial systems, when examined in detail, are nonlinear, to some extent. For such

systems, the issue of designing nonlinear observers in the presence of unknown input

and output is of a practical importance.

Fault detection and isolation (FDI) is but only one example of how important

such observers can be when used for the purpose of process monitoring in presence of

unknown disturbances. Thus, in the design of observers for fault detection purposes,

the observers should be insensitive to unknown disturbances in the inputs and outputs

and at the same time perform the task of monitoring and state estimating accurately.

Page 2 of 21
Preprints submitted to IJC Aldeen & Sharma

Whereas for linear systems, necessary and sufficient conditions for existence of

unknown-input observers (UIO) have been fully established (Kudva et al. 1980), little

work has been reported in the field of nonlinear unknown input observer (NUIO)

theory. It is for this reason that recent research efforts have been directed into

addressing this deficiency.

One of the earliest studies in this field is the work of (Wunnenberg 1990). In this

approach, the nonlinearities acting on the system are assumed to be functions of the

inputs and available measurable outputs. However, such assumption places restriction

on its applicability. (Seliger and Frank 1991) proposed an alternative approach for a

general class of nonlinear systems. The approach provides existence conditions and

proposes a design method based on disturbance decoupling through state

transformation. A draw back of this approach is that convergence of the error

dynamics is not proven or discussed. In (Corless and Tu 1998), the problem of state

and input estimation is considered for a class of uncertain systems composed of a

linear part and an uncertain/nonlinear/time-varying part. The approach incorporates

the main results of (Kudva et al. 1980) and (Park and Stein 1988) and is applicable to

a larger class of systems. Another notable contribution is due to (Koenig and Mammar

2001) reported an unknown input reduced order observer for nonlinear systems for the

purpose of FDI, which is also based upon complete disturbance decoupling. A more

recent contribution in this field is reported by (Navarro, Banks et al 2003) where an

observer design for general class of nonlinear systems using a novel linear

approximation is reported. (Sharma and Aldeen 2004) extended the work in (Koenig

and Mammar 2001) by considering the problem of simultaneously estimating the

states and unknown inputs and/or outputs using the same nonlinear observer.

Page 3 of 21
Preprints submitted to IJC Aldeen & Sharma

In another approach by (Shields 1997), an unknown input observer is designed

for nonlinear descriptor systems that can be used in residual generation for fault

detection. The advantage of this approach is that robustness to unknown inputs can be

investigated without resorting to disturbance decoupling.

None of the above mentioned approaches consider the problem of

simultaneously estimating directly unknown inputs and faults in the plant (actuator

faults) and measurements (sensor faults). Therefore the main feature of the proposed

technique is the use of a nonlinear observer that can provide asymptotic state

estimation, unknown input estimation, and simultaneously detect and identify a

number of faults associated with the plant and the measurement. As a result, the

problem of redundancy associated with fault detection by residual generation can be

significantly lessened.

2. SYSTEM DESCRIPTION AND ANALYSIS

Consider the following nonlinear system

x = hx ( x, u , f e ) (1)

y = hy ( x, u, f w ) (2)

If the nonlinear function hx ( x, u , f ) is differentiable with respect to x , then this class

of system may be expressed in terms of a linear unforced part and a nonlinear state

dependent controlled part as follows (Rajamani 1998)

x = Ax + Bu + Ef e + g ( x, u ) (3)

Page 4 of 21
Preprints submitted to IJC Aldeen & Sharma

y = Cx + Du + Wf w (4)

where, x ∈ℜn ; u ∈ ℜ p ; y ∈ ℜm ; f e ∈ℜqe ; f w ∈ℜqw are the state, input, output,

unknown input and output disturbances, respectively, and matrices A, B, E , C , D and

W are constant and of appropriate dimensions.

Remark 2.1: Matrices E and W are assumed to be of full rank, otherwise the following

rank is carried out. Assume that rank { E} = qe < m . Then we can arrange E so that

⎡f ⎤
Ef e = [ E1 E2 ] ⎢ e1 ⎥ = E1 f e1 + E2 f e 2 and rank { E} = rank { E1} = qe . This implies that
⎣ fe2 ⎦

E2 may be expressed in terms of E1 as E2 = E1 ⋅ K , where K contains as many

columns as there are rows in E1 . Thus, we can write E1 f e1 + E1 Kf e 2 = E1 ( f e1 + Kf e 2 )

= E1 f e ' . It follows that if we excite the observer with two signals, namely f e1 and f e ' ,

( )
then we can construct the unknown signal fˆe 2 from fˆe 2 = K + fˆe' − fˆe1 . Because K

has full column rank, fˆe 2 has a unique solution. The same procedure can be used in

the case where the output fault matrix W is degenerate.

In the ensuing analysis, following have been assumed to hold:

Assumption 1: The nonlinearity g ( x, u ) is assumed to be Lipschitz (Isidori 1995,

Khalil 2002), with Lipschitz constant γ 0 . That is,

g ( z1 , u ) − g ( z2 , u ) ≤ γ 0 z1 − z2 ∀z1 , z2 ∈ℜn (5)

Assumption 2: m > qe + qw .

Remark 2.2: Assumption 2 may be relaxed if for physical reasons, it may not be

satisfied. The relaxation, however, can be affected by designing more than one

observer, each for a subset of the faults signals. If, however, due to some practical and

Page 5 of 21
Preprints submitted to IJC Aldeen & Sharma

accessibility constraints either of this cannot be done, then the technique will have to

be abandoned.

3. Development of Unknown Input Observer Based Fault Detection Scheme

In this section, we introduce an unknown input observer based fault detection

scheme suitable for multivariable systems. The scheme is based on the idea of

decoupling the state and output equations into fault-free and fault-dependent parts.

The fault-free part is, then, used to design an observer that would guarantee

estimation of the entire state vector irrespective of the magnitude and nature of any

fault signals. Once this task is achieved, then the observer is used to provide estimates

of the state vector to a fault detection filter, which is designed specifically to provide

estimates of all fault and unknown input signals. This development is outlined next.

3.1 Decoupling the output equation

Using singular value decomposition, matrix W can be expressed as:

⎡∑ ⎤
W = U ∑W V T = [U1 U 2 ] ⎢ 2 ⎥ V T = U1 ∑ 2 V T (6)
⎣0⎦

where U ∈ℜm×m , ∑W ∈ℜm×qw , V T ∈ℜqw ×qw ,U1 ∈ℜm×qw ,U 2 ∈ℜ m×( m − qw ) and ∑ 2 ∈ℜqw ×qw .

Equation (3) may then be written as:

y = Cx + Du + U ∑W V T fW (7)

Define f w = V T f w ; f w ∈ ℜq and pre-multiply equation (7) by U T to obtain:


w

⎡U1T y ⎤ ⎡ y1 ⎤
U y=⎢ T ⎥⎢ ⎥
T
(8)
⎣U 2 y ⎦ ⎣ y2 ⎦

Page 6 of 21
Preprints submitted to IJC Aldeen & Sharma

⎡ D1 ⎤ ( m − qw )× p
Define U D  ⎢ ⎥ , D1 ∈ ℜ w×p
and D2 ∈ ℜ
T q
, then equation (7) becomes
⎣ 2⎦
D

⎡ y1 ⎤ ⎡U1T ⎤ ⎡ D1 ⎤ ⎡∑2 ⎤
⎢ y ⎥ = ⎢ T ⎥ Cx + ⎢ D ⎥ u + ⎢ 0 ⎥ f w (9)
⎣ 2 ⎦ ⎣U 2 ⎦ ⎣ 2⎦ ⎣ ⎦

From (9) we have:

y1 = U1T Cx + D1u + ∑ 2 f w (10)

y2 = U 2T Cx + D2u = U 2T y (11)

Thus the output equation has now been decoupled, as shown in equations (10) and

(11). Note that equation (11) reveals that y2 is not affected by the output fault

signal f w .

3.2 Decoupling the state equation

Singular value decomposition of matrix E ∈ℜn×qe gives:

⎡∑ ⎤
E = Q ∑ E RT = [Q1 Q2 ] ⎢ 1 ⎥ RT = Q1 ∑1 RT (12)
⎣0⎦

where, Q ∈ℜn×n , ∑ ∈ℜn×qe , R ∈ℜqe ×qe , Q1 ∈ℜ n×qe , Q2 ∈ℜ n×( n − qe ) and ∑1 ∈ℜqe ×qe

⎡Q ⎤ ⎡ g ( x, u ) ⎤
Q T g ( x, u ) = ⎢ 1 ⎥ g ( x, u ) = ⎢ 1 ⎥ = G ( x, u ) .
⎣Q2 ⎦ ⎣ g 2 ( x, u ) ⎦

Define: f e = RT f e ; f e ∈ℜqe and

⎡ x1 ⎤
⎢x ⎥ = x  Q x
T
(13)
⎣ 2⎦

so that

Page 7 of 21
Preprints submitted to IJC Aldeen & Sharma

⎡A A12 ⎤ T ⎡ B1 ⎤
A  QT AQ = ⎢ 11 ⎥ ;Q B = B  ⎢ ⎥ (14)
⎣ A21 A22 ⎦ ⎣ B2 ⎦

then equation (3) becomes:

x1 = A11 x1 + A12 x2 + B1u + Σ E1 f e + g1 ( x, u ) (15)

x2 = A21 x1 + A22 x2 + B2u + g 2 ( x, u ) (16)

⎡x ⎤
Use this coordinate transformation x = Qx = Q ⎢ 1 ⎥ in equation (11) to obtain
⎣ x2 ⎦

y2 = Cx + D2u (17)

where C ∈ℜ(
m − qw )×n
is defined as:

C  U 2T CQ = ⎡⎣C1 C2 ⎤⎦ (18)

with C1 ∈ ℜ(
m − qw )× qe
, and C2 ∈ ℜ(
m − qw )×( n − qe )
. Then equation (17) becomes:

y2 = C1 x1 + C2 x2 + D2u (19)

Since, rank C1 = qe , an ( m − qw ) × ( m − qw ) non-singular matrix may be constructed

from:

⎡C1+ ⎤
N =⎢ ⎥ (20)
⎣M ⎦

is the pseudo-inverse of C1 , defined as C1+ = ( C1T C1 ) C1T , and


qe ×( m − qw ) −1
where, C1+ ∈ ℜ

M ∈ ℜ(
m − qw − qe )×( m − qw )
is an arbitrarily selected matrix so that N ∈ ℜ(
m − qw )×( m − qw )
is

nonsingular. Premultiplying (17) by (20) gives:

⎡C1+ ⎤ ⎡C1+ ⎤ ⎡ x1 ⎤ ⎡C1+ ⎤


⎢ ⎥ y2 = ⎢ ⎥ ⎡⎣C1 C2 ⎤⎦ ⎢ ⎥ + ⎢ ⎥ D2u (21)
⎣M ⎦ ⎣M ⎦ ⎣ x2 ⎦ ⎣ M ⎦

Page 8 of 21
Preprints submitted to IJC Aldeen & Sharma

Since C1+ C1 = I qe , equation (21) yields:

x1 = C1+ ( y2 − D2u − C2 x2 ) (22)

My2 = MC1 x1 + MC2 x2 + MD2u (23)

Substituting (22) into (16) yields:

x2 = ( A22 − A21C1+ C2 ) x2 + ( B2 − A21C1+ D2 ) u + A21C1+ y2 + g 2 ( x, u ) (24)

Using y2 = U 2T y in (24) gives

x2 = A 2 x2 + B 2u + G 2 y + g 2 ( x, u ) (25)

where, A 2 = ( A22 − A21C1+ C2 ) , B 2 = ( B2 − A21C1+ D2 ) and G 2 = A21C1+U 2T . Using

equation (20) in equation (23) gives:

( ) ( ) (
M I m − q2 − C1C1+ y2 = M I m − q2 − C1C1+ C2 x2 + M I m − q2 − C1C1+ D2u ) (26)

Define:

( ) ( ) (
C  M I m − qw − C1C1+ C2 ; H  M I m − qw − C1C1+ U 2T ; D  M I m − qw − C1C1+ D2 )
(27)

Using definitions in equation (26) gives:

Hy  + Du
 = Cx  (28)
2

Define:

y 2  Hy  + Du
 = Cx  (29)
2

Thus we now have a fault-free system described by the state equation (25) and the

output equation (29), i.e.

Page 9 of 21
Preprints submitted to IJC Aldeen & Sharma

x2 = A 2 x2 + B 2u + G 2 y + g 2 ( x, u ) (30)

 + Du
y 2 = Cx 
2 (31)

where, the dimension of the state vector, x2 , is n − qe .

3.3 Observer Design

{ }
If the pair A 2 , C is observable, then the following system can act as an observer

for the system described by equations (30):


(
xˆ2 = A 2 xˆ2 + B 2u + G 2 y + g 2 ( xˆ , u ) + L y 2 − Cx
 ˆ − Du
2
 ) (32)

where L is the observer gain matrix and must be found so that the observer matrix

A 2 − LC is stable.

Once an estimate of x2 is obtained from the observer described in (32), the

estimate of x1 can be found from equation (22) as

(
xˆ1 = C1+ U 2T y − D2u − C2 xˆ2 ) (33)

Finally, the estimate, x̂ , of the original overall state vector, x , is now constructed

from equation (13) as:

⎡ xˆ ⎤
xˆ = Qxˆ = Q ⎢ 1 ⎥ (34)
⎣⎢ xˆ2 ⎦⎥

3.4 Estimation of Fault Signals

Page 10 of 21
Preprints submitted to IJC Aldeen & Sharma

In this section, we derive two fault estimation schemes, one for fault signals that

appear in the state equation and the other for the fault signals that appear in the output

equation.

3.4.1 Estimation of output fault signals

Using equation (33) in equation (34) gives the following:

xˆ = Qx xˆ2 + Qy y + Qu u (35)

⎡ −C + C ⎤ ⎡C +U T ⎤ ⎡ −C + D ⎤
where, Qx  Q ⎢ 1 2 ⎥ , Qy  Q ⎢ 1 2 ⎥ , and Qu  Q ⎢ 1 2 ⎥ . From equation (10),
⎣ I ⎦ ⎣ 0 ⎦ ⎣ 0 ⎦

and after replacing x by its estimate, x̂ , we have:

fˆw = V Σ −21U1T ( y − Cxˆ − Du ) (36)

3.4.1 Estimation of input fault signals

From equation (35) we have


xˆ = Qx xˆ2 + Qy y + Qu u (37)

Using equation (32) in equation (37) gives

⎣( ) ( ) ( )
xˆ = Qx ⎡ A 2 − LC xˆ2 + B 2 − LD u + G 2 + LH y ⎤ + Qy y + Qu u

(38)

From equations (15) and (38), the estimate of the fault signal f e may be obtained,

after replacing x by its estimate, x̂ , as:

(
fˆe = RΣ1−1Q1T Ψ1 xˆ2 + Ψ 2 y + Ψ 3 y + Ψ 4u + Ψ 5 g ( xˆ , u ) + Ψ 6u ) (39)

Page 11 of 21
Preprints submitted to IJC Aldeen & Sharma

( ) (
where, Ψ1 = Qx A 2 − LC − AQx , Ψ 2 = Qx G 2 + LH − AQy , Ψ 3 = Qy , )
( )
Ψ 4 = Qx B 2 − LD − AQu − B , Ψ 5 = QxQ2 and Ψ 6 = Qu .

4. STABILITY ANALYSIS

Let us define the following state error equations

e = x − xˆ (40)

and e2 = x2 − xˆ2 (41)

Then from equations (30)-(32), we have

( )
e2 = A 2 − LC e2 + Q2 g (42)

where, g = g ( x, u ) − g ( xˆ , u ).

Let us introduce the following Lyapunov function

V = e2T Pe2 (43)

This implies

( ) (
V = e2T ⎡ P A 2 − LC + A 2 − LC ) P ⎤ e2 + 2e2T PQ2 g
T
(44)
⎢⎣ ⎥⎦

( )
If the pair A 2 , C is detectable, then by appropriate choice of gain L, stability of

( A − LC )
2 can be ensured. Therefore, by Lyapunov stability theory there exists a

positive definite symmetric matrix X such that

( ) (
P A 2 − LC + A 2 − LC )
T
P = −X (45)

Using equation (45) in (44) yields

V = −e2T Xe2 + 2e2T PQ2 g (46)

Further, using assumption (5), we obtain

Page 12 of 21
Preprints submitted to IJC Aldeen & Sharma

V ≤ − X e2 + 2 e2 PQ2γ 0 e
2
(47)

From equations (34) and (35) we have

( )
e = x − xˆ = Qx x2 − xˆ2 = Qx e2 (48)

Thus, inequality (47) is expressed as

V ≤ − X e2 + 2γ 0 PQ2Qx e2
2 2

≤ − e2 ⎡⎣σ min ( X ) − 2γ 0σ max ( PQ2Qx ) ⎤⎦


2
(49)

where, σ max and σ min represent maximum and minimum singular values,

respectively.

Therefore, asymptotic stability of error dynamics is guaranteed if the Lipschitz

condition satisfies the following inequality:

σ min ( X )
γ0 < (50)
2σ max ( PQ2Qx )

5. NUMERICAL EXAMPLES AND SIMULATION RESULTS

In order to illustrate the methodology described in sections 2-4, we consider two

different examples of nonlinear systems.

Example 1: Train System

In this example, the design technique presented in the paper is applied for estimating

unknown states and detecting faults in the train system shown in figure (1).

Page 13 of 21
Preprints submitted to IJC Aldeen & Sharma

z3 z1

k, ρ u
MC ME
spring

Carriage Engine
Figure 1: A train system.

This system comprises an engine, E, and a carriage, C, coupled by a spring with a

stiffness coefficient k representing the linear characteristics and ρ representing the

nonlinear characteristics. The engine mass is represented by M E , while the carriage

mass is represented by M C . The train is moved by the engine traction force, u, which

may be lost for some practical reasons. The loss of traction may be modelled as a

fault. A nonlinear state space model describing the train motion incorporating possible

loss of traction is derived as follows

⎡ 0 1 0 0 ⎤ ⎡ 0 ⎤
⎢ k ⎥ ⎡ 0 ⎤ ⎡ 1 ⎤ ⎢ ρ ⎥
⎢− −α g
k
0 ⎥ ⎢ 1 ⎥ ⎢ 1 ⎥ ⎢ ( z2 + z4 ) ⎥
2

⎢ ME ME ⎥ ⎢ ⎥ ⎢− ⎥ ⎢ME ⎥
z = ⎢ ⎥ z + ⎢ME ⎥u + ⎢ ME ⎥ f + ⎢ ⎥
⎢ 0 0 0 1 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 0 ⎥
⎢ k k ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ ρ 2⎥
⎢ 0 − −α g ⎥ ⎢⎣ 0 ⎥⎦ ⎢⎣ 0 ⎥⎦ ⎢ ( z2 + z4 ) ⎥
⎣ MC MC ⎦ ⎣ MC ⎦

⎡1 0 0 0 ⎤
y=⎢ ⎥z
⎣0 1 0 0 ⎦

Page 14 of 21
Preprints submitted to IJC Aldeen & Sharma

where α represents the rolling friction and f represents loss of traction (fault). The

state vector is defined as z = [ z1 z4 ] , where z1 and z2 are position and


T
z2 z3

velocity of the engine while z3 and z4 are those of the carriage. For simulation

purpose, the following set of data is used: M E = 10 Kg , M C = 5 Kg , k = 4.87 N/s ,

α = 0.5 s/m , ρ = 1 N/s and g = 9.8 m / s 2 .

Observer design: Using Assumption 1, the Lipschitz constant is computed

as γ 0 = 0.088 . More specifically, we calculated the Lipschitz constant as the

supremum of the magnitude of the partial derivative of the nonlinear term f ( z ) . That

is,

∂f ( z )
γ0 =
∂z z = z ( 0)

where z (0) represents the equilibrium point of the nonlinear train system under

consideration.

(
In addition, the following value of observer gain is chosen to ensure that A 2 − LC is )
Hurwitz:

⎡ 0.7816 ⎤
L = ⎢⎢ 0.1174 ⎥⎥ .
⎢⎣ 0.1846 ⎥⎦

( )
This observer gain matrix results in eig A 2 − LC = ( −0.7816, − 0.98 ± j 0.1166 ) .

Furthermore, choosing X = I and solving (45) for P gives

⎡0.6397 0.096 0.1501 ⎤


P = ⎢⎢ 0.096 1.6165 −0.4887 ⎥⎥
⎢⎣ 0.1501 −0.4887 0.5584 ⎥⎦

Page 15 of 21
Preprints submitted to IJC Aldeen & Sharma

For the X and P above, the RHS of condition (50) is computed as:

σ min ( X )
= 0.5529 > γ 0 = 0.088
2σ max ( PQ2Qx )

This ensures asymptotic stability of the error dynamics, i.e. asymptotic convergence

of the nonlinear observer to the true states of the system.

Fault filter design: The fault detection filter described in equation (39) has the

following design parameters:

(
fˆe = RΣ1−1Q1T Ψ1 xˆ2 + Ψ 2 y + Ψ 3 y + Ψ 4u + Ψ 5 g ( xˆ , u ) )

⎡ −0.3749 0 0⎤ ⎡0.3749 0 ⎤ ⎡0 0⎤
⎢ 0.4870 −0.4870 0 ⎥⎥ ⎢ 0 1.96 ⎥⎥ ⎢0 1 ⎥⎥
Ψ1 = ⎢ ; Ψ2 = ⎢ ; Ψ3 = ⎢ ;
⎢ −0.0118 0 0⎥ ⎢ 0.0118 0 ⎥ ⎢0 0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ −0.0564 0 0⎦ ⎣0.0564 0 ⎦ ⎣0 0⎦

⎡1 0 0 0⎤
⎢0 0 0 0 ⎥⎥
Ψ 4 = [ 0 0 0 0] ; Ψ 5 = ⎢
T
.
⎢0 0 1 0⎥
⎢ ⎥
⎣0 0 0 1⎦

Simulation results: To examine the performance of the designed observer and fault

filter, the following simulation study is carried out. The system is assumed to be in

equilibrium, when at time t = 0 (s) the observer is switched on with zˆ(0) ≠ z (0) . At

time t = 10 (s) the train increases traction gradually until it reaches full traction at

time t = 15 (s) ; this can be modelled as a unit ramp applied over a period of

10 ≤ t ≤ 15 ( s) . Thereafter the traction is held constant until an engine fault develops

at t = 35 (s) resulting in gradual loss of traction. At time t = 40 (s) complete loss of

traction occurs. The engine fault is then gradually cleared over the period

Page 16 of 21
Preprints submitted to IJC Aldeen & Sharma

55 ≤ t ≤ 60 ( s) when at t = 60 (s) full traction is restored. The simulation results for

this case are shown in figures (2a-2c). It is clear from all of these figures that the

observer converges to the true state of the system in about 10s, which means that the

fault filter is now able to detect occurrence of faults.

Figure (2a), shows the response of the fault detection filter. As expected, the

filter does not respond to the application of the traction at t = 10 (s) . However when

the engine loses its traction at time t = 35 (s) , it starts to respond. As shown in the

figure the output of the fault filter increases gradually, flattens and then gradually

declines, image mirroring the fault signal over the fault period 35 ≤ t ≤ 60 ( s) .

Figures (2b-2c) exhibit the properties of the nonlinear unknown input observer

by accurately reconstructing the unmeasured states z1 and z4 accurately despite the

aforementioned changes in the engine traction force, u .

Figure 2(a): Estimation of Fault signal

Page 17 of 21
Preprints submitted to IJC Aldeen & Sharma

Figure 2(b): Estimation of state z1

Figure2(c): Estimation of state z4

6. CONCLUSION

A new scheme for the design of nonlinear observer and fault detection filters for

a class of nonlinear systems is presented in this paper. It is shown that the scheme is

able to reconstruct the occurrence of faults and unknown input and output

disturbances. A stability analysis is carried out using a Lyapunov function. Finally,

the effectiveness of the technique is illustrated with the help of two numerical

examples.

7. REFERENCES

Page 18 of 21
Preprints submitted to IJC Aldeen & Sharma

M. Corless and J. Tu, “State and Input Estimation for a Class of Uncertain systems”,

Automatica, 34, pp 757-764, 1998.

Y. Guan and M. Saif, “A novel approach to the design of unknown input observers”,

IEEE Transactions on Automatic Control, 36, pp. 632-635, 1998.

Hou, M. and P.C. Muller, “Design for observers for linear systems with unknown

inputs”, IEEE Transactions on Automatic Control, 37, pp. 871-874, 1992.

M. Hou and P.C. Muller, “Disturbance Decoupled Observer Design: A Unified

Viewpoint”, IEEE Transactions on Automatic Control, 39, pp. 1338-341, 1994.

A. Isidori, Nonlinear control systems, New York: Springer, 1995.

Hassan K. Khalil, Nonlinear systems, Upper saddle River, N.J.: Prentice Hall, 2002.

D. Koenig and S. Mammar, “Design of a class of reduced order unknown inputs

nonlinear observer for fault diagnosis”, in Proceedings of the American control

Conference, Arlington, VA, 2001, pp. 4353-4358.

D. Koenig and R.J. Patton, “New design of robust Kalman filters for fault detection

and isolation”, in Proceedings of the 14th World Congress. International

Federation of Automatic Control, Elsevier, 17, 1999, pp 247-252.

Page 19 of 21
Preprints submitted to IJC Aldeen & Sharma

J.E. Kurek, “The state vector reconstruction for linear systems with unknown inputs”,

IEEE transactions on Automatic control, 25, pp. 1120-1122, 1983.

J.E. Kurek, “Reconstruction of states of linear multivariable systems with unknown

inputs”, International Journal of control, 37, pp. 1459-1463, 1983.

P. Kudva, N. Viswanadham and A. Ramakrishna, “Observer design for linear systems

with unknown inputs”, IEEE Transactions on Automatic Control, 25, pp. 113-

115, 1980.

M.S. Park and J. L. Stein, “Closed-loop state and input observer for systems with

unknown inputs”, International Journal of Control, 48, pp. 1121-1136, 1988.

S. Raghavan and J.K. Hedrick, “Observer design for a class of nonlinear systems”,

International Journal of Control, 59, pp. 515-28, 1994.

R. Rajamani, “Observers for Lipschitz nonlinear systems”, IEEE Transactions on

Automatic Control, 43, pp. 397-400, 1998.

R. Seliger and P.M. Frank, “Fault-Diagnosis by disturbance decoupled nonlinear

observers”, in Proceedings of the 30th IEEE conference on Decision and

Control, Brighton, England, 1991, pp. 2248-2253.

R. Sharma and M. Aldeen, “Estimation of unknown disturbances in nonlinear


systems”, in Control 2004, Bath, UK.

Page 20 of 21
Preprints submitted to IJC Aldeen & Sharma

D.N. Shields, “Observer design and detection for nonlinear descriptor systems”,

International Journal of Control 67, pp.153-68, 1997.

J. Wunnenberg, Observer based Fault Detection in Dynamic Systems, VDI -

Fortschrittsber., Chapter 4, VDI- Verlag, Reihe 8, Nr.222. Dusseldorf, Germany,

pp 64-72, (1990).

F. Yang and R.W. Wilde, “Observers for linear systems with unknown inputs”, IEEE

Transactions on automatic control, 33, pp. 677-681, 1988.

Page 21 of 21

You might also like