Fault detection and diagnosis and

unknown input reconstruction based on

parity equations concept
Sumislawska, Malgorzata
Unpublished PhD thesis deposited in CURVE February 2015

Original citation:
Sumislawska, M. (2012) Fault detection and diagnosis and unknown input reconstruction
based on parity equations concept. Unpublished PhD thesis. Coventry: Coventry University

Copyright © and Moral Rights are retained by the author(s) and/ or other copyright
owners. A copy can be downloaded for personal non-commercial research or study,
without prior permission or charge. This item cannot be reproduced or quoted extensively
from without first obtaining permission in writing from the copyright holder(s). The
content must not be changed in any way or sold commercially in any format or medium
without the formal permission of the copyright holders.

CURVE is the Institutional Repository for Coventry University

http://curve.coventry.ac.uk/open
 Fault detection and diagnosis
and unknown input
reconstruction based on parity
equations concept

Malgorzata Sumislawska
MSc Teleinformatics, MSc Control Engineering

October 2012

A thesis submitted in partial fulfilment of the University’s requirements

for the degree of Doctor of Philosophy.

Control Theory and Applications Centre

Coventry University
Abstract

There are two main threads of this thesis, namely, an unknown (unmeasurable) input
reconstruction and fault detection and diagnosis. The developed methods are in the
form of parity equations, i.e. finite impulse response filters of the available input and
output measurements.
In the first thread the design of parity equations for the purpose of an unknown
input reconstruction of linear, time-invariant, discrete-time, stochastic systems is taken
into consideration. An underlying assumption is that both measurable system inputs
as well as the outputs can be subjected to noise, which leads to an errors-in-variables
framework. The main contribution of the scheme is accommodation of the Lagrange
multiplier method in order to minimise the influence of the noise on the unknown
input estimate. Two potential applications of the novel input reconstruction method
are proposed, which are a control enhancement of a hot strip steel rolling mill and an
estimation of a pollutant level in a river.
Furthermore, initial research is conducted in the field of the unknown input recon-
struction for a class of nonlinear systems, namely, Hammerstein-Wiener systems, where
a linear dynamic block is preceded and followed by a static nonlinear function. Many
man-made as well as naturally occurring systems can be accurately described using
Hammerstein-Wiener models. However, it is considered that not much attention has
been paid to Hammerstein-Wiener systems in the errors-in-variables framework and in
this thesis it is aimed to narrow this gap.
The second thread considers a problem of robust (disturbance decoupled) fault de-
tection as well as fault isolation and identification. Unmeasurable external stimuli,
parameter variations or discrepancies between the system and the model act as distur-
bances, which can obstruct the fault detection process and lead to false alarms. Thus,
a fault detection filter needs to be decoupled from the disturbances. In this thesis
the right eigenstructure assignment method used for the robust fault detection filter
design is extended to systems with unstable invariant zeros. Another contribution re-
gards the design of robust parity equations of any arbitrary order using both left and
right eigenstructure assignment. Furthermore, a parity equation-based fault isolation
and identification filter is designed which provides an estimate of the fault. A simple
method for the calculation of thresholds whose violation indicates a fault occurrence is
also proposed for the errors-in-variables framework.

i
Acknowledgements

Firstly, I would like to express my gratitude towards my Director of Studies, Prof

Keith J. Burnham and my supervisor Dr Tomasz Larkowski for their support and
encouragement during this project. I would like to acknowledge Dr Leszek Koszalka
and Dr Iwona Poźniak-Koszalka for enabling me the initial opportunity to study at the
Control Theory and Applications Centre. I would also like to give my special thanks
to Ivan Zajı́c, Prof Peter Young, Dr Gerald Hearns, Dr Peter Reeve, and Prof Ron J.
Patton for inspiring discussions and their valuable comments.

ii
Contents

Page

Abstract i

Acknowledgements ii

Contents iii

Nomenclature vii
Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

List of Algorithms xiii

1 Introduction, motivation and outline of approach 1

1.1 Introduction and motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Unknown input reconstruction . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Fault detection and diagnosis . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Outline of approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.2 Outlines of chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Review 9
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Linear system representation . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Polynomial representation . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 State-space representation . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.3 Polynomial and state-space representations of stochastic systems . 10
2.2.4 Invariant zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.5 Properties of a linear system in geometric theory . . . . . . . . . . 12
2.3 Block oriented models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Unknown input reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . 18

iii
 CONTENTS

2.4.1 MVU state estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.4.2 Input estimation (INPEST) method . . . . . . . . . . . . . . . . . 21
2.4.3 Efficacy measure of compared algorithms . . . . . . . . . . . . . . . 22
2.5 Introduction to fault detection and diagnosis . . . . . . . . . . . . . . . . . 22
2.5.1 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5.2 Fault detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.6 Robust fault detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.6.1 Robust fault detection via right eigenstructure assignment . . . . 27
2.6.2 Disturbance decoupling in geometric approach . . . . . . . . . . . . 28
2.6.3 Robust fault detection via left eigenstructure assignment . . . . . 29
2.6.4 Design of first order PE using left eigenstructure assignment . . . 31
2.7 Fault isolation and identification . . . . . . . . . . . . . . . . . . . . . . . . 32
2.7.1 Fault isolation via diagnostic observers . . . . . . . . . . . . . . . . 32
2.7.2 Geometric properties of fault isolation filter . . . . . . . . . . . . . 36
2.7.3 Fault identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.8 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3 Parity equations-based unknown input reconstruction for linear stochas-

tic systems 38
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2 Linear system representation . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 Design of unknown input reconstructor . . . . . . . . . . . . . . . . . . . . 43
3.3.1 Parity equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.2 Unknown input estimation . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3.3 Selection of optimal W . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3.4 Design of PE-UIO for OE systems . . . . . . . . . . . . . . . . . . . 49
3.4 Analysis in frequency domain . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Two stage PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.1 Two stage filter design . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.2 Analysis in frequency domain . . . . . . . . . . . . . . . . . . . . . . 58
3.6 Generalised two stage PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.7 Tutorial examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.7.1 PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.7.2 Two stage PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.8 Comparison with other methods . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.9 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4 Parity equations-based unknown input reconstruction for Hammerstein-

Wiener systems 78
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

iv
 CONTENTS

4.3 PE-UIO for Hammerstein-Wiener systems . . . . . . . . . . . . . . . . . . . 82

4.3.1 Parity relations for Hammerstein-Wiener system . . . . . . . . . . 82
4.3.2 Unknown input estimation . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3.3 Confidence bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.3.4 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Adaptive order PE-UIO for Hammerstein-Wiener systems . . . . . . . . . 95
4.4.1 Choice of s(t) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4.2 Variable estimation lag . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4.3 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.5 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5 Robust fault detection via eigenstructure assignment 105

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.2 Robust fault detection via right eigenstructure assignment for systems
with unstable invariant zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.2.1 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.2.2 Solution for q = 1 with a single invariant zero . . . . . . . . . . . . . 111
5.2.3 Solution for q = 1 with multiple invariant zeros . . . . . . . . . . . . 115
5.2.4 General solution for q ≥ 1 . . . . . . . . . . . . . . . . . . . . . . . . 118
5.2.5 Consideration of residual response to fault . . . . . . . . . . . . . . 120
5.2.6 Differences and similarities with fault isolation filter of Chen and
Speyer (2006a) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.2.7 Tutorial examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.3 Design of robust parity equations using right eigenstructure assignment . 133
5.3.1 Finite-time convergent observer . . . . . . . . . . . . . . . . . . . . . 134
5.3.2 Proposed scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.3.3 Design of robust PE . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.3.4 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.4 Design of robust parity equations using left eigenstructure assignment . . 142
5.4.1 Design of robust PE . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.4.2 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.5 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

6 Fault isolation via diagonal PE 151

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3 Design of fault isolation filter . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.4 Fault identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.4.1 Change of coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.4.2 Steady state gain calculation . . . . . . . . . . . . . . . . . . . . . . 157
6.5 Consideration of measurement noise . . . . . . . . . . . . . . . . . . . . . . 159

v
 CONTENTS

6.6 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

6.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

7 Potential applications 165

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.2 Steel rolling mill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.2.1 Description of the plant . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.2.2 Plant model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.2.3 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
7.3 Hydrological application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.4 Critical appraisal of practical application of developed methods . . . . . . 174

8 Conclusions & further work 176

8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
8.1.1 Unknown input reconstruction . . . . . . . . . . . . . . . . . . . . . 176
8.1.2 Fault detection and diagnosis . . . . . . . . . . . . . . . . . . . . . . 178
8.1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
8.2 Further work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

References 180

Appendices 193

A Calculation of parameters xij 194

B Proof of Lemma 5.4 198

C Demonstration of Remark 6.3 201

vi
Nomenclature

Abbreviations

AO-PE-UIO-HW adaptive order parity equations-based unknown input observer for Hammerstein-
Wiener systems
ARMAX . . . . . . . autoregressive model with moving average and exogenous input
ARX . . . . . . . . . . . autoregressive model with exogenous input
cf. . . . . . . . . . . . . . . confer (compare)
DC . . . . . . . . . . . . . direct current
DRFDF . . . . . . . . deadbeat robust fault detection filter
e.g. . . . . . . . . . . . . . exempli gratia (for example)
EIV . . . . . . . . . . . . errors-in-variables
INPEST . . . . . . . . input estimation (input reconstruction method)
LMI . . . . . . . . . . . . linear matrix inequalities
MIMO . . . . . . . . . . multiple-input multiple-output
MVU . . . . . . . . . . . minimum variance unbiased (MVU) state and input estimator
NMSS . . . . . . . . . . non-minimum state-space
OE . . . . . . . . . . . . . output error model
PE . . . . . . . . . . . . . parity equations
PE-UIO . . . . . . . . parity equations-based unknown input observer
PE-UIO-HW . . . parity equations-based unknown input observer for Hammerstein-Wiener systems
PIP . . . . . . . . . . . . proportional integral plus
PIP-LQ . . . . . . . . proportional integral plus linear quadratic
RFDF . . . . . . . . . . robust fault detection filter
SISO . . . . . . . . . . . single-input single-output
SVF . . . . . . . . . . . . state variable feedback

Notation
Latin variables

ai . . . . . . . . . . . . . . autoregressive parameter in polynomial model

A ............... state transition matrix in state-space model
Ac1 , Ac2 . . . . . . . . filter state transition matrices
Ap . . . . . . . . . . . . . piston area (in steel rolling mill model)
(i)
Aλ , A∗e , A∗e , Aw auxiliary matrices
˜ Ã, A′ . . . . . . . . .
Ã, auxiliary matrices

b(t) . . . . . . . . . . . . auxiliary vector

vii
 Nomenclature

bi . . . . . . . . . . . . . . exogenous parameter in polynomial model

B .............. input matrix of known input in state-space model
˜ ..............
B̃ auxiliary matrix

ci . . . . . . . . . . . . . . moving average parameter in ARMAX model

C .............. output matrix in state-space model or compensation variable in steel rolling mill
model
C′ . . . . . . . . . . . . . . auxiliary matrix

d(t) . . . . . . . . . . . . disturbance signal

d1 , d2 , d2 . . . . . . . . damping coefficients (in steel rolling mill model)
di (t) . . . . . . . . . . . ith element of d(t)
d∗ (t) . . . . . . . . . . . d(t), whose elements, di (t), are delayed, respectively by δi
D .............. feedforward matrix of known input in state-space model

e(t) . . . . . . . . . . . . noise term

e ............... matrix of directions of elements of d(t)
ei . . . . . . . . . . . . . . direction of di (t)
ē . . . . . . . . . . . . . . . matrix built from matrices ēi
ēi . . . . . . . . . . . . . . matrix whose image is sum of image of ei and images of invariant zero directions
of (A, ei , C)
E .............. input matrix of disturbance signal in state-space model
Ei . . . . . . . . . . . . . . ith column of input matrix of disturbance signal in state-space model

f (⋅) . . . . . . . . . . . . function to be minimised by the Lagrange multiplier method

f ............... matrix of directions of elements of µ(t)
fi . . . . . . . . . . . . . . direction of µi
f¯ . . . . . . . . . . . . . . . matrix built from matrices f¯i
f¯i . . . . . . . . . . . . . . matrix whose image is sum of image of fi and images of invariant zero directions
of (A, fi , C)
F .............. input matrix of fault signal in state-space model
Fc (t) . . . . . . . . . . . Coulomb friction (in steel rolling mill model)
Ff ric (t) . . . . . . . . total friction force (in steel rolling mill model)
Fh (t) . . . . . . . . . . . hydraulic force (in steel rolling mill model)
Fi . . . . . . . . . . . . . . ith column of F
Fr (t) . . . . . . . . . . . roll force (in steel rolling mill model)
Fs (t) . . . . . . . . . . . Stribeck friction (stiction) in steel rolling mill model
Fv (t) . . . . . . . . . . . viscous friction (in steel rolling mill model)

g(⋅) . . . . . . . . . . . . . constraint function in the Lagrange multiplier method

g, gi . . . . . . . . . . . . auxiliary scalar
G .............. input matrix of unknown input in state-space model
G1 , G2 . . . . . . . . . . controller gains (in steel rolling mill model)
G′ . . . . . . . . . . . . . . auxiliary matrix
Gu (z) . . . . . . . . . . z-domain transfer function between u0 (t) and y(t)
Gv (z) . . . . . . . . . . z-domain transfer function between v(t) and y(t)
G′v (z) . . . . . . . . . . auxiliary transfer function

h(t) . . . . . . . . . . . . exit gauge (in steel rolling mill model)

ĥ(t) . . . . . . . . . . . . estimate of exit gauge (in steel rolling mill model)

viii
 Nomenclature

href (t) . . . . . . . . . exit gauge reference signal (in steel rolling mill model)
H(t) . . . . . . . . . . . input gauge (in steel rolling mill model)
H .............. feedforward matrix of unknown input in state-space model
H′ . . . . . . . . . . . . . . auxiliary matrix

J, J1 , J2 . . . . . . . . . gain matrices

k1 , k2 . . . . . . . . . . . spring constants (in steel rolling mill model)

k3 . . . . . . . . . . . . . . steel strip spring constant (in steel rolling mill model)
K(t) . . . . . . . . . . . gain matrix (used by MVU)
K, K1 , K2 , K ′ . . . gain matrices
Kc . . . . . . . . . . . . . hydraulic oil compressibility coefficient (in steel rolling mill model)
Kp . . . . . . . . . . . . . proportional gain of hydraulic piston controller (in steel rolling mill model)

l ................ stroke length of the piston (in steel rolling mill model)
lj , lj∗ . . . . . . . . . . . . transposes of left eigenvectors of filter state transition matrix

m .............. number of system outputs

m1 . . . . . . . . . . . . . mass of hydraulic piston (in steel rolling mill model)
m2 . . . . . . . . . . . . . mass of backup roll (in steel rolling mill model)
m3 . . . . . . . . . . . . . mass of work roll (in steel rolling mill model)
M .............. mill modulus (spring constant) in steel rolling mill model
M̂ . . . . . . . . . . . . . . estimated mill modulus (in steel rolling mill model)
Mv (z) . . . . . . . . . . auxiliary transfer function
M (t) . . . . . . . . . . . auxiliary matrix (used by MVU) or stacked vector of last τ + 1 values of µ(t)
(depending on context)

n ............... order of the system

na . . . . . . . . . . . . . . order of autoregressive polynomial
nb . . . . . . . . . . . . . . order of exogenous polynomial
nc . . . . . . . . . . . . . . order of moving average polynomial
Nv (z) . . . . . . . . . . auxiliary transfer function

p ............... number of known inputs to a system

p(t) . . . . . . . . . . . . hydraulic pressure (in steel rolling mill model)
pi . . . . . . . . . . . . . . element of vector P
P .............. auxiliary vector or matrix
P (λi ) . . . . . . . . . . auxiliary function of eigenvalue λi
Pl,k . . . . . . . . . . . . auxiliary term
P x (t),P d (t) . . . . . submatrices of state and input estimation error covariance matrix
P dx (t),P xd (t) . . . submatrices of state and input estimation error covariance matrix

q ............... number of disturbance signals

qf (t) . . . . . . . . . . . fluid flow to hydraulic capsule (in steel rolling mill model)
Q .............. block Toeplitz matrix or gain matrix (depending on context)
Q̃ . . . . . . . . . . . . . . covariance matrix of ξ(t)

r ............... number of fault signals

r(t) . . . . . . . . . . . . residual
ri (t) . . . . . . . . . . . . ith element of r(t)

ix
 Nomenclature

rq . . . . . . . . . . . . . . number of rows of Q
R .............. covariance matrix of ζ(t) or an auxiliary matrix (depending on context)
R̃ . . . . . . . . . . . . . . covariance matrix (used in MVU)

s ............... parity space order

S ............... auxiliary matrix

T ............... block Toeplitz matrix or similarity transformation matrix (depending on context)

T1 , T2 . . . . . . . . . . . submatrices of the similarity transformation matrix T
T′ .............. auxiliary matrix

u0 (t) . . . . . . . . . . . noise-free known input

u(t) . . . . . . . . . . . . measured input
ũ(t) . . . . . . . . . . . . input measurement noise
ũ∗ (t) . . . . . . . . . . . auxiliary variable
ū0 (t) . . . . . . . . . . . input to linear dynamic block in Hammerstein-Wiener system
˜(t) . . . . . . . . . . . .
ū estimation error of ū0 (t)
U0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of u0 (t) (used in parity equations)
U0 (z) . . . . . . . . . . u0 (t) in z-domain
U (t) . . . . . . . . . . . . stacked vector of last s + 1 values of u(t) (used in parity equations)
U (z) . . . . . . . . . . . u(t) in z-domain
Ũ (t) . . . . . . . . . . . . stacked vector of last s + 1 values of ũ(t) (used in parity equations)
Ũ (z) . . . . . . . . . . . ũ(t) in z-domain
Ũ ∗ (t) . . . . . . . . . . . stacked vector of last s + 1 values of ũ∗ (t) (used in parity equations)
Ũ ∗ (z) . . . . . . . . . . ũ∗ (t) in z-domain
˜ (t) . . . . . . . . . . . .
Ū ˜(t)
stacked vector of last s + 1 values of ū

v(t) . . . . . . . . . . . . unknown (unmeasurable) input

vc , v1 , v2 . . . . . . . . auxiliary friction coefficients (in steel rolling mill model)
(i)
v, vj , vj . . . . . . . . auxiliary vectors
Ve . . . . . . . . . . . . . . matrix whose columns are eigenvectors of filter state transition matrix
V (t) . . . . . . . . . . . . stacked vector of last s + 1 values of unknown input (used in parity equations)
V (z) . . . . . . . . . . . v(t) in z-domain
v̂(t) . . . . . . . . . . . . unknown input estimate
V̂ (z) . . . . . . . . . . . v̂(t) in z-domain
v ′ (t) . . . . . . . . . . . . auxiliary variable
V ′ (t) . . . . . . . . . . . stacked vector of last s + 1 values of v ′ (t) (used in parity equations)
V ′ (z) . . . . . . . . . . . v ′ (t) in z-domain

w qi , w ξ i . . . . . . . . . auxiliary polynomial parameters

(i) ′(i)
wj , wj , wj ... right eigenvectors of filter state transition matrix
∗
wj . . . . . . . . . . . . . auxiliary vector
W .............. vector, which belongs to Γ– or auxiliary matrix (depensing on context)
Wu , Wy . . . . . . . . . parity matrices
W (z) . . . . . . . . . . . polynomial of z-variable defined by appropriate elements of W
WQ (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of W Q
WT (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of W T
WT ′ (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of W T ′
WΞ (z) . . . . . . . . . . polynomial of z-variable defined by appropriate elements of W Ξ

x
 Nomenclature

x(t) . . . . . . . . . . . . state vector instate space model

x̂(t) . . . . . . . . . . . . state estimate
xi,j . . . . . . . . . . . . . auxiliary scalars
X .............. auxiliary matrix

y0 (t) . . . . . . . . . . . noise-free output in output-error case

y(t) . . . . . . . . . . . . measured output
ỹ(t) . . . . . . . . . . . . output measurement noise
ȳ0 (t) . . . . . . . . . . . output of linear dynamic block
ȳ˜(t) . . . . . . . . . . . . estimation error of ȳ0 (t)
Ȳ0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of ȳ0 (t)
Y (t) . . . . . . . . . . . . stacked vector of last s + 1 values of measured output (used in parity equations)
Ȳ˜ (t) . . . . . . . . . . . . stacked vector of last s + 1 values of ȳ˜(t)
Y0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of noise-free output (used in parity equations)
Ỹ (t) . . . . . . . . . . . . stacked vector of last s + 1 values of ỹ(t)
Y (z) . . . . . . . . . . . y(t) in z-domain

z(t) . . . . . . . . . . . . position of the hydraulic piston (in steel rolling mill model)
zi . . . . . . . . . . . . . . system zero
zi (t) . . . . . . . . . . . . state estimate
zref (t) . . . . . . . . . hydraulic piston position reference signal (in steel rolling mill model)

Greek variables

αi , αi′ . . . . . . . . . . . . . . . . . . . . . auxiliary parameters

βi . . . . . . . . . . . . . . . . . . . . . . . . . ith diagonal element of Σ in multiple output OE case or auxiliary pa-
rameter
χ ......................... auxiliary vector
δ .......................... system delay
δi . . . . . . . . . . . . . . . . . . . . . . . . . auxiliary term
ǫ(t) . . . . . . . . . . . . . . . . . . . . . . . auxiliary noise term
ǫ∗ (t) . . . . . . . . . . . . . . . . . . . . . . auxiliary noise term
φu (t), φy (t) . . . . . . . . . . . . . . . auxiliary variance terms
ϕ(⋅) . . . . . . . . . . . . . . . . . . . . . . . Hammerstein nonlinearity
γ ......................... row vector of Γ–
Γ ......................... extended observability matrix
Γ– . . . . . . . . . . . . . . . . . . . . . . . . left nullspace of Γ
η(⋅) . . . . . . . . . . . . . . . . . . . . . . . Wiener nonlinearity
η −1 (⋅) . . . . . . . . . . . . . . . . . . . . . inverse of η(⋯)
κ ......................... auxiliary scalar
λ ......................... Lagrange multiplier
(i)
λj , λj . . . . . . . . . . . . . . . . . . . . eigenvalue of filter state transition matrix
Λe . . . . . . . . . . . . . . . . . . . . . . . . diagonal matrix whose diagonal elements are eigenvalues of filter state
transition matrix
µ(t) . . . . . . . . . . . . . . . . . . . . . . fault signal
µc . . . . . . . . . . . . . . . . . . . . . . . . Coulomb friction level (in steel rolling mill model)
µi (t) . . . . . . . . . . . . . . . . . . . . . . ith element of fault signal
µs . . . . . . . . . . . . . . . . . . . . . . . . Stribeck friction coefficient (in steel rolling mill model)
µv . . . . . . . . . . . . . . . . . . . . . . . . viscous friction coefficient (in steel rolling mill model)
µ̂(t) . . . . . . . . . . . . . . . . . . . . . . estimate of fault signal

xi
 Nomenclature

ν ......................... input derivative weighting

Θ(i) , Θ̄(i) . . . . . . . . . . . . . . . . . . auxiliary matrices
Π ......................... input matrix of noise term in state-space model
Ω ......................... feedforward matrix of noise term in state-space model or set of all
invariant zeros of (A, e, C) depending on context
Ωi . . . . . . . . . . . . . . . . . . . . . . . . set of all invariant zeros of (A, ei , C)
Σ,Σe , Σũ , Σũe . . . . . . . . . . . . covariance matrices
Σũ , Σµ̄ . . . . . . . . . . . . . . . . . . . . covariance matrices
τ ......................... unknown input estimation lag, convergence time of finite time-
convergent state observer
ξ(t) . . . . . . . . . . . . . . . . . . . . . . . process noise vector in state space model
Ξ ......................... block Toeplitz matrix
ψ ......................... auxiliary vector
Ψ ......................... auxiliary matrix
Ξ(t) . . . . . . . . . . . . . . . . . . . . . . auxiliary matrix
ζ(t) . . . . . . . . . . . . . . . . . . . . . . . output noise vector in state space model

Operators and symbols

∈ ............... element in
∪ ............... union of two sets or subspaces
/ ............... difference of two sets or subspaces
∩ ............... intersection of two sets or subspaces
⊂ ............... a subset of
⊆ ............... a subset or equal to
⊕ ............... direct sum
Rm×n . . . . . . . . . . . m × n dimensional space of real numbers
arg min f (x) . . . . value of x that minimises f (x)
rank(M ) . . . . . . . rank of matrix M
E {⋅} . . . . . . . . . . . . expected value operator
MT . . . . . . . . . . . . transpose of matrix M
M −1 . . . . . . . . . . . . inverse of matrix M
M ............. Moore-Penrose pseudo inverse of matrix M
M– . . . . . . . . . . . . . nullspace of matrix M
V– . . . . . . . . . . . . . . orthogonal completion of subspace V
Im{A} . . . . . . . . . . image of A
Ker{A} . . . . . . . . . kernel of A
round(q) . . . . . . . rounding of the scalar q to the nearest natural number
span{A} . . . . . . . . subspace spanned by A
sumrow (A) . . . . . column vector whose elements are sums of the appropriate rows A
var(e(t)) . . . . . . . variance of e(t)
z −1 . . . . . . . . . . . . . backward shift operator

xii
List of Algorithms

2.1 MVU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1 PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 PE-UIO for single-output OE . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3 PE-UIO for multiple-output OE . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4 Two stage PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.5 Generalised two stage PE-UIO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.1 PE-UIO-HW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 AO-PE-UIO-HW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.1 RFDF via right eigenstructure assignment, q = 1, q1 = 1 . . . . . . . . . . . . . 114
5.2 RFDF via right eigenstructure assignment, q = 1, q1 ≥ 1 . . . . . . . . . . . . . 117
5.3 RFDF via right eigenstructure assignment, q ≥ 1, qi ≥ 0 . . . . . . . . . . . . . 119
5.4 Robust PE via right eigenstructure assignment . . . . . . . . . . . . . . . . . . 139
5.5 Robust PE via left eigenstructure assignment . . . . . . . . . . . . . . . . . . . 143
6.1 Fault isolation using directional PE . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.2 Fault isolation and identification via diagonal PE . . . . . . . . . . . . . . . . 159

xiii
Chapter 1

Introduction, motivation and

outline of approach

1.1 Introduction and motivation

Safety considerations provided an essential objective for motivating the research on
fault detection and diagnosis. As air travel became popular, and plans of the first
manned space missions were made, it became crucial to ensure safety of the people
on board in the case of malfunction. It was important to isolate a fault and handle
it accurately. Early analogue methods for fault detection in DC power systems for
aircrafts have been presented in (Kaufmann & Finison 1952). Then, a decade later an
automated take-off monitoring system for an aircraft was patented by Craddock (1962).
1960s were times of the first space missions, which boosted research on methods for
fault detection and location of faults (Janis 1963, Mast, Mayper & Pilnick 1966). In the
1970s fault accommodation in a spacecraft (i.e. reconfiguration of the control system
such that safety of operation can be achieved in the presence of a fault) has been
proposed (Kennedy 1970).
Another field where fault diagnosis is crucial for the safety of operators is a nuclear
reactor, for which safety control apparatus has been developed, among others, in (Dever
1960). Garrick, Gekler, Goldfisher, Karcher, Shimizu & Wilson (1967) proposed a fault
detection method using logic gates for a nuclear reactor. Gradually, as computational
power became more available, application of fault diagnosis to other industrial processes
has been considered. Lee (1962) proposed a method that can locate faults in a 555-
transistor digital system. Halton (1963) proposed an automated checkout of a drone.
(In the 1960’s the term ‘checkout’ has been widely used for fault diagnosis.) A design
of software for an automated checkout has been summarised in (Jirauch 1967).
A breakthrough in the fault detection and diagnosis field, which formed a basis for
the modern fault detection and diagnosis, was use of the Luenberger state observer
(Luenberger 1964) for the purpose of fault diagnosis by Beard (1971) and Jones (1973).

1
 1. Introduction, motivation and outline of approach

Fault diagnosis schemes based on the so-called Beard-Jones fault detection filter have
been in use up to this day in various industrial processes, e.g. chemical batch reac-
tors (Pierria, Paviglianiti, Caccavale & Mattei 2008), satellite attitude control systems
(Wang, Jiang & Shi 2008), or gas turbines (Gao, Breikin & Wang 2007).
Another important method which is used up to this day are parity equations (PE),
developed independently by Mironovski (1979) and Chow & Willsky (1984). Chan,
Hua & Hong-Yue (2006) applied PE for fault diagnosis of DC motors. A recent work
of Berriri, Naouar & Slama-Belkhodja (2011) presents a parity space approach for
diagnosis of a current sensor electrical system.
Fault detection and diagnosis algorithms are used practically in every industry
(Isermann 2005). They are not only used to detect abrupt malfunctions, but also to
signal wear and tear of machine parts, hence indicating when particular parts should
be replaced and facilitate the maintenance process.
The other topic, which is explored in this thesis, is an unknown (unmeasurable)
input reconstruction. Early contributions to this subject can be found in (Dorato 1969,
Sain & Massey 1969, Moylan 1977). Approximate input reconstruction has been used in
(Fu, Yan, Santillo, Palanthandalam-Madapusi & Bernstein 2009, Fu, Kirtikar, Zattoni,
Palanthandalam-Madapusi & Bernstein 2009) for diagnosing aircraft control surfaces.
Rocha-Cozatl, Moreno & Vande Wouwer (2012) utilised a continuous-discrete unknown
input observer in order estimate unknown variables in phytoplanktonic cultures. Recent
work of Czop (2011) presents reconstruction of the passenger vehicle wheel vertical
movement under ride conditions.

1.2 Problem statement

1.2.1 Unknown input reconstruction
A system is a real world entity, which can be represented using a mathematical model
that describes relationships between system inputs and outputs. Inputs may not only
represent external stimuli, but also discrepancies between the model and the real sys-
tem. Some of the input signals can be measured, some, however, are inaccessible for
measurement. A representation of a system with unknown (unmeasurable) inputs is de-
picted in Fig. 1.1. The aim of the unknown input reconstruction problem is to estimate
the unmeasurable inputs to the system based on the known accessible measurements.
Methods considered in this thesis assume that the mathematical model of the system
is known.

1.2.2 Fault detection and diagnosis

A fault means a malfunction of a system and/or component. A fault detection and
diagnosis process can be divided into three stages, which are defined by answers to the
following questions:

2
 1. Introduction, motivation and outline of approach

unmeasured
inputs

?
measured measured
inputs outputs
System

Figure 1.1: Representation of a system with unknown inputs

1. Is there a malfunction in the system? If the answer is ‘yes’, a fault has been
detected. Thus, the first stage of the process is called fault detection.

2. Which component is faulty? Determining, which component is malfunctioning,

is referred to as fault isolation.

3. By how much the component is faulty? Determining the magnitude of the fault is
denoted as fault identification, i.e. defining the quantity by which the particular
system parameter deviated from its acceptable/nominal value.

In the literature fault isolation and identification are often denoted as fault diagnosis.
Faults are usually modelled as extra inputs to the system. Thus, a fault diagnosis
process can be understood as unknown input estimation.

1.3 Outline of approach

1.3.1 Methodology
Algorithms developed in this thesis are built on existing schemes, by extending/adapting
them, by combining two (or more) different methods or by applying a well known
scheme for different purposes than originally designed. Proposed algorithms are de-
scribed in details in a form that allows their straightforward implementation using
computer software, e.g. Matlab. For a better understanding of the devised schemes tu-
torial examples are presented and care is taken to ensure reproducibility of the results.
A benchmark comparison of some of the proposed algorithms with examples from the
literature is also provided where possible.
A review of well known methods existing in the literature, on which the schemes
proposed in this thesis are built, is given in the review Chapter 2. Furthermore, Chap-
ters 3–6 start with a short literature review on the particular topic that each of the
chapters explores. All abbreviations and nomenclature used throughout this thesis are
given in a separate section at the beginning of the thesis. Some of the terms have
different meanings depending on the context in which they are used. Therefore, for

3
 1. Introduction, motivation and outline of approach

completeness, at the beginning of Chapters 2–7 the nomenclature used in the par-
ticular chapter is provided. Additionally, Chapters 3–7 start with an indication of a
preliminary reading from specific sections of this thesis. Outlines of following chapters
are provided in Subsection 1.3.2.

1.3.2 Outlines of chapters

Chapter 2: The aim of this chapter is to review available methods for unknown in-
put reconstruction and fault detection and diagnosis, which are the bases for
algorithms developed in further chapters. Firstly, the notation for the system
representation used throughout this thesis is provided. Then, unknown input
reconstruction methods are reviewed and two schemes are presented, which are
further used as benchmarks for the novel scheme developed in Chapter 3. Fur-
thermore, the nomenclature used for fault detection and diagnosis is provided and
well known methods for robust fault detection and isolation are presented, which
forms a basis for development of algorithms in Chapters 5–6.

Chapter 3: In this chapter a method for unknown (unmeasurable) input reconstruc-

tion is proposed, i.e. parity equation-based unknown input observer (PE-UIO).
The algorithm is devised for systems that are subjected to process and measure-
ment noise in the errors-in-variables (EIV) framework, i.e. the known input is
affected by white, Gaussian, zero-mean independent and identically distributed
(i.i.d.) noise sequences, whereas the output is subjected to coloured noise. PE
are used for the purpose of an unknown input reconstruction with a Lagrange
multiplier method utilised to find an optimal solution minimising the effect of
noise on the unknown input estimate. The order of the parity space is a tuning
parameter which allows adjustment of the bandwidth and, hence, noise filtering
properties of the filter. The efficacy of the novel scheme is compared with those of
two known methods, namely, minimum variance unbiased (MVU) state and input
estimator, see (Gillijns & De Moor 2007b), and input estimation (INPEST), see
(Young & Sumislawska 2012).

Chapter 4: This chapter builds on the algorithm developed in Chapter 3. The un-
known input reconstruction method is extended to a class of nonlinear systems,
namely Wiener-Hammerstein systems, where a linear dynamic block is preceded
and followed by a memoryless nonlinear function. Similarly, as in Chapter 3,
an EIV framework is considered. Due to nonlinearities, the impact of the noise
on the unknown input estimate depends on the values of the known input and
output themselves, and is changing over time. Therefore, an adaptation scheme
is devised, which allows adjustment of the order of the parity space (and, con-
sequently, the filter bandwidth) based on the change of the measured input and
output.

4
 1. Introduction, motivation and outline of approach

Chapter 5: There are two main outcomes of this chapter: firstly, the robust fault
detection filter based on right eigenstructure is extended to systems with unstable
invariant zeros, which extends the applicability of the aforementioned scheme. It
is also demonstrated that the devised algorithm is computationally simpler than
that of Chen & Speyer (2006a). Then, a robust PE of user-defined order is
designed using right and left eigenstructure assignment. In order to obtain an
open-loop solution (i.e. equivalent to PE) a finite time convergent state observer
is utilised. The disturbance decoupling property of the novel scheme is proven
algebraically and its efficacy is shown using a numerical example.

Chapter 6: This chapter builds on Chapter 5. Decoupling properties of the robust

PE designed in Chapter 5 are used to devise a fault isolation and identification
filter, which generates an estimate of the fault signal.

Chapter 7: Practical applications of the algorithms developed in Chapter 3 are pro-

posed. The PE-based unknown input reconstruction scheme is used to improve
control performance of a simulated single stand of a steel rolling mill. Further-
more, it is proposed to apply the unknown input reconstruction algorithm to a
hydrological application.

Chapter 8: In this chapter concluding remarks are given and proposals for further
work are stated.

A structural representation of the flow of developments carried out in this thesis is

depicted in Fig. 1.2.

1.4 Contributions
Contributions of the author are listed in descending order with respect to their consid-
ered relative significance.

1. Parity equation-based unknown input reconstruction for linear stochastic systems:

The main contribution of the scheme is use of the Lagrange multiplier method to
find optimal filter parameters such that the effect of noise on the unknown input
estimate is minimised. The parity equation-based unknown input observer (PE-
UIO) has been originally developed in [1]1 for systems with a single output in an
output error (OE) case (i.e. when the output of the system is subjected to white,
Gaussian, zero-mean noise). Then, in [2] and [3], the method has been extended
for systems in the EIV framework, i.e. when both input and output are affected by
white, Gaussian, zero-mean, and mutually uncorrelated noise sequences. Then,
in [4], the scheme has been extended to a multivariable case.
1
Note that the references in square brackets are the publications of the author.

5
 1. Introduction, motivation and outline of approach

Chapter 2
Review

Chapter 3 Chapter 5
Parity equations−based Robust fault detection
unknown input reconstruction via eigenstructure assignment
for linear stochastic systems

Chapter 4 Chapter 6
Parity equations−based Fault isolation
unknown input reconstruction via diagonal PE
for Hammerstein−Wiener systems

Chapter 7
Potential applications

Figure 1.2: Structural representation of a logical flow of developments of this thesis

[1] Sumislawska, M., Burnham, K. J., Larkowski, T., Design of unknown input
estimator of a linear system based on parity equations. In Proc. of the
17th International Conference on Systems Science, pages 81–90, Wroclaw,
Poland, September 2010
[2] Sumislawska, M., Larkowski, T., Burnham, K. J., Design of unknown input
reconstruction algorithm in presence of measurement noise. In Proc. of
the 8th European ACD2010 Workshop on Advanced Control and Diagnosis,
pages 213–216, Ferrara, Italy, November 2010
[3] Sumislawska, M., Larkowski, T., Burnham, K. J., Design of unknown input
reconstruction filter based on parity equations for EIV case. In Proc. of the
18th IFAC World Congress, pages 4272–4277, Milan, Italy, 2011
[4] Sumislawska, M., Larkowski, T., Burnham, K. J., Parity equations-based
unknown input estimator for multiple-input multiple-output linear systems.
Systems Science, 36(3):49–56, 2010

2. Extension to of the PE-UIO to Wiener-Hammerstein systems: The PE-UIO has

6
 1. Introduction, motivation and outline of approach

been extended to a class of nonlinear systems, where the system consists of a

linear dynamic block preceded and followed by nonlinear static functions.

[5] Sumislawska, M., Larkowski, T. and Burnham, K. J., Unknown input recon-
struction observer for Hammerstein-Wiener systems in the errors-in-variables
framework. In Proc. of the 16th IFAC Symposium on System Identification,
pages 1377–1382, Brussels, Belgium, 2012.

3. Enhancement of the control performance of a steel rolling mill: The control of a

single finishing stand in a steel rolling mill is affected by a friction force. Strongly
nonlinear properties of the friction force and its positive feedback in the control
loop lead to oscillations which affect the quality of the final product. It is proposed
here to estimate the friction force using PE and feed it back to the controller. As
a result of the compensation, significant reduction of the amplitude of oscillations
has been demonstrated in a simulation study.

[6] Sumislawska, M., Burnham, K. J., Hearns, G., Larkowski, T., Reeve, P.
J., Parity equation-based friction compensation applied to a rolling mill. In
Proc. of the UKACC International Conference on Control, pages 1043–1048,
Coventry, UK, September 2010

4. Robust fault detection based on eigenstructure assignment: Chen & Patton (1999)
have shown that a special case of the robust fault detection filter based on left
eigenstructure assignment is equivalent to first order PE. Also up to date the link
between the right eigenstructure assignment-based robust fault detection filter
and PE has not been derived. The main contribution of [7] and [8] is the design
of robust PE of an arbitrary order using, respectively, right and left eigenstructure
assignment.

[7] Sumislawska, M., Larkowski, T., Burnham, K. J., Design of parity equations
using right eigenstructure assignment. In Proc. of the 21st International
Conference on Systems Engineering, pages 367–370, Las Vegas, USA, August
2011
[8] Sumislawska, M., Larkowski, T. and Burnham, K. J., Design of robust par-
ity equations of user-defined order using left eigenstructure assignment. In
Proc. of the 9th European Workshop on Advanced Control and Diagnosis,
Budapest, Hungary, November 2011

5. Extension of the right eigenstructure assignment method to systems with unsta-

ble invariant zeros: Invariant zeros of the residual response to disturbances are
unobservable modes of the robust fault detection filter via right eigenstructure
assignment, i.e. the invariant zeros of the system become poles of the fault de-
tection filter. Therefore, design of a stable filter becomes impossible when those

7
 1. Introduction, motivation and outline of approach

zeros are unstable. A solution is proposed in this thesis which allows the design
of stable robust fault detection filters for systems with unstable invariant zeros.

6. Fault isolation and identification based on diagonal PE using right eigenstructure

assignment: This development builds on contributions 4 and 5 above. The right
eigenstructure is used to design a PE, whose output is an estimate of the fault
vector. A system with multiple faults is considered. The scheme is applicable to
systems, whose response to faults contains unstable invariant zeros.

8
Chapter 2

Review

2.1 Introduction
The purpose of this chapter is to familiarise the reader with the notation and the
background knowledge used to develop algorithms in the next chapters. Firstly, the
notation of a linear discrete-time time-invariant stochastic system representation and a
class of nonlinear systems, namely, block-oriented systems, is provided in, respectively,
Section 2.2 and Section 2.3. In Section 2.4 the problem of an unknown input recon-
struction is presented. Two methods for a reconstruction of the unknown input signal
are presented, which are further used for a benchmark comparison with the algorithms
developed in Chapter 3. Section 2.5 provides the nomenclature used in the subject of
fault diagnosis as well as describes open- and closed-loop fault detection. Furthermore,
in Section 2.6 the problem of a robust (disturbance decoupled) fault detection is re-
viewed. A closely related fault isolation and identification is presented in Section 2.7.
Concluding remarks are given in Section 2.8.

2.2 Linear system representation

The algorithms presented in this thesis are designed for discrete-time time-invariant
systems. The schemes proposed in Chapters 3, 5, and 6 are derived for linear sys-
tems, whilst the algorithms developed in Chapter 4 are devised for a class of nonlinear
systems. The algorithms presented in this thesis utilise a state-space representation
of a linear system, see Subsection 2.2.2. The unknown input reconstruction schemes
developed in Chapters 3 and 4 are designed for stochastic systems, i.e. those which
are subjected to random noise. The notation of a stochastic linear discrete-time time-
invariant model is defined in Subsection 2.2.3. State-space representations of known
polynomial noise models, see (Ljung 1999), are presented as well. For completeness, a
polynomial representation of a linear system is defined in Subsection 2.2.1.

9
 2. Review

2.2.1 Polynomial representation

Assume that a linear dynamic discrete-time time-invariant multiple-input multiple-
output (MIMO) system with p inputs and m outputs is represented by an nth order
state-space equation of the following form (Ljung 1999):

na nb
y(t) = − ∑ ai y(t − i) + ∑ bj u(t − j) (2.1)
i=1 j=0

The terms u(t) ∈ Rp and y(t) ∈ Rm refer to, respectively, the input and output vectors,
whilst na and nb , with na ≥ nb , are the orders of the auto-regressive and exogenous
parameters, respectively, and ai ∈ Rm×m and bi ∈ Rm×p are coefficient matrices.

2.2.2 State-space representation

The algorithms, which are developed within the framework of this thesis utilise the
state-space form of a linear system, which is given by:

x(t + 1) = Ax(t) + Bu(t)

(2.2)
y(t) = Cx(t) + Du(t)

where x(t) ∈ Rn is the state vector, whilst A ∈ Rn×n , B ∈ Rn×p , C ∈ Rm×n , and D ∈ Rm×p .
A notation (A, B, C, D) is used to refer to the system (2.2). In the case when D = 0,
the system (2.2) is denoted as (A, B, C). System (2.1) can be described by an observer
canonical form of a state-space model (Ljung 1999, Yiua & Wang 2007), where matrices
A, B, C, and D are:

⎡ ⎤
⎢ b1 − a 1 b0 ⎥
⎢ ⎥
⎡ −a
⎢ 0 ⋯ 0 ⎤
⎥
⎢
⎢ b2 − a 2 b0 ⎥
⎥
⎥ ⎢ ⎥
I
⎢
0 I ⋯ 0 ⎥ ⎢ ⎥
1
⎢ −a
⎢ ⎥ ⎢ ⋮ ⎥
⎥ ⎢ ⎥
2
⎢
⎢
A=⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥
⎥ B=⎢
⎢ bnb − anb b0 ⎥
⎥ C = [ I 0 ⋯ 0 ] D = b0 (2.3)
⎢ ⎥ ⎢ ⎥
⎢ −ana −1
⎢ 0 0 ⋯ I ⎥
⎥
⎢
⎢ −anb +1 b0 ⎥⎥
⎢ ⎥ ⎢ ⎥
⎢ −ana
⎣ 0 0 ⋯ 0 ⎥
⎦
⎢
⎢ ⋮ ⎥
⎥
⎢ ⎥
⎢ −ana b0 ⎥
⎣ ⎦

2.2.3 Polynomial and state-space representations of stochastic sys-

tems
Consider system (2.2) affected by stochastic noise. A general representation of the
system is given by the following set of equations:

x(t + 1) = Ax(t) + Bu0 (t) + Πe(t)

y(t) = Cx(t) + Du0 (t) + Ωe(t) (2.4)
u(t) = u0 (t) + ũ(t)

10
 2. Review

where Π ∈ Rn×m , Ω ∈ Rm×m . The terms u0 (t) ∈ Rp and y(t) ∈ Rm refer to, respectively,
the input and output vectors. The term e(t) ∈ Rm is a column vector of m zero-mean,
white, Gaussian, independent and identically distributed (i.i.d.) noise sequences. The
term ũ(t) ∈ Rp is a vector of white, zero-mean, Gaussian i.i.d. noise sequences, which is
uncorrelated with e(t). Equation (2.4) is a generalised representation of a linear system
and can be simplified in more specific cases, some of which are given below.

Auto-regressive model with moving average and exogenous input (ARMAX)

A MIMO auto-regressive model with a moving average and exogenous input (ARMAX)
is given by, see (Ljung 1999, Yiua & Wang 2007):

na nb nc
y(t) = − ∑ ai y(t − i) + ∑ bj u(t − j) + ∑ ck e(t − k) (2.5)
i=1 j=0 k=0

where u(t) and y(t) are, respectively, the input and output vectors of the system and
e(t) is a vector of white, zero-mean, Gaussian, i.i.d. noise sequences. The terms na ,
nb , nc , with na ≥ nb and na ≥ nc , are the orders of the auto-regressive, exogenous and
moving average parameters, respectively, and ai ∈ Rm×m , bi ∈ Rm×p and ci ∈ Rm×m are
coefficient matrices. The last component of the right-hand side of (2.5) refers to the
moving average (coloured) process noise of the system.
The state-space system matrices (2.2) for the ARMAX model (2.5) in the observer
canonical form are given by:

⎡ ⎤ ⎡ ⎤
⎢ b1 − a 1 b0 ⎥ ⎢ c 1 − a1 c 0 ⎥
⎢ ⎥ ⎢ ⎥
⎡ −a 0 ⋯ 0 ⎤ ⎢ ⎥ ⎢ c 2 − a2 c 0 ⎥
⎢ ⎥ ⎢ b2 − a 2 b0 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
I
⎢ −a 0 I ⋯ 0 ⎥ ⎢ ⎥ ⎢ ⎥
1

⎢ ⎥ ⎢ ⋮ ⎥ ⎢ ⋮ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
2
A=⎢
⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥
⎥ B=⎢
⎢
⎥
bnb − anb b0 ⎥ Π = ⎢
⎢ c nc − a nc c 0 ⎥
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ −ana −1 0 0 ⋯ I ⎥ ⎢ ⎥ ⎢ −anc +1 c0 ⎥
(2.6)
⎢ ⎥ ⎢ −anb +1 b0 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ −ana 0 0 ⋯ 0 ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎢ ⋮ ⎥ ⎢ ⋮ ⎥
⎢ ⎥ ⎢ ⎥
⎢ −ana b0 ⎥ ⎢ −ana c0 ⎥
⎣ ⎦ ⎣ ⎦
C =[ I 0 ⋯ 0 ] D = b0 Ω = c0

The ARMAX model assumes that the input u(t) is known exactly (there is no noise
present on the input variable), hence ũ(t) = 0 and u(t) = u0 (t). Note that an autore-
gressive model with an exogenous input (ARX) is obtained from an ARMAX model by
setting ci , i = 1, ⋯, nc , to zero.

Output error (OE) model

An OE model assumes, that there is no process noise present in the system, however
the noise-free output y0 (t) is subjected to zero-mean, white, Gaussian measurement

11
 2. Review

noise e(t), see (Ljung 1999):

na nb
y0 (t) = − ∑ ai y0 (t − i) + ∑ bj u(t − j)
i=1 j=0 (2.7)
y(t) = y0 (t) + e(t)

This case can be modelled by the system representation (2.2), where matrices A, B, C,
and D are all given as in the ARMAX case. The matrix Π is null, and Ω is diagonal.
Also, there is no noise present on the input variable, hence ũ(t) = 0.

Errors-in-variables (EIV) framework

In the EIV framework, see, for example, (Söderström 2007), all measured variables, i.e.
the inputs and outputs of the system, are affected by zero-mean, white, Gaussian, i.i.d.
measurement noise sequences. This can be represented by (2.2), where ũ(t) ≠ 0, Π = 0,
and Ω is diagonal.

2.2.4 Invariant zeros

An invariant zero of the system (A, B, C, D) is such a value zi for which the Rosenbrock
system matrix, defined as:

⎡ ⎤
⎢ zi I − A −B ⎥
⎢
P (zi ) = ⎢ ⎥
⎥
⎢
(2.8)
−D ⎥
⎣ ⎦
C

loses its rank (MacFarlane & Karcanias 1976). Two vectors are associated with an
invariant zero: the invariant zero state direction v ∈ Rn and the invariant zero input
direction g ∈ Rp , which conform the following equation (El-Ghezawi, Billings & Zinober
1983, Patel 1985, Patel & Munro 1982):

⎡ ⎤⎡ ⎤
⎢ zi I − A −B ⎥ ⎢ v ⎥
⎢ ⎥⎢ ⎥ = 0
⎢ ⎥⎢ ⎥
⎢ −D ⎥ ⎢ g ⎥
(2.9)
⎣ ⎦⎣ ⎦
C

2.2.5 Properties of a linear system in geometric theory

Robust fault detection and fault isolation may be easier to understand if the reader
is familiar with the basics of a geometric approach. The aim of this subsection is to
demonstrate some geometric properties of linear systems.
Consider a matrix A ∈ Rn×m . An image of A is defined as a set of all vectors Ax for
any arbitrary x ∈ Rm :
Im{A} = {Ax ∶ x ∈ Rm } (2.10)

An image of A is sometimes defined as span{A}, i.e. the subspace spanned by A, which

12
 2. Review

is a set of all possible linear combinations of columns of A, i.e.

m
span{A} = {∑ αi Ai ∶ αi ∈ R} (2.11)
i=1

where Ai , i = 1, ⋯, m are columns of A and αi are arbitrary scalars. The kernel of A is

a set of all x ∈ Rm for which Ax = 0, i.e.

Ker{A} = {x ∈ Rm ∶ Ax = 0} (2.12)

Denote an n-dimensional space over the field of real numbers as X . Consider a matrix
V ∈ Rn×k , where k ≤ n. Denote V = Im{V }; then V is a k-dimensional subspace of the
space X . (Note that X = Im{I}.) Consider the following subspaces V, Y, and Z of
vector spaces Rn and Rm and a matrix A ∈ Rm×n . For completeness, basic operations
on subspaces are given below (Halmos 1958, Basile & Marro 2002):

1. Sum:
Z = V + Y ∶= {z ∶ z = v + y, v ∈ V, y ∈ Y} (2.13)

2. Intersection:
Z = V ∩ Y ∶= {z ∶ z ∈ V, z ∈ Y} (2.14)

3. Direct sum:

Z = V ⊕ Y ∶= {z ∶ z = v + y, v ∈ V, y ∈ Y, V ∩ Y = 0} (2.15)

4. Linear transformation:

Y = AV ∶= {y ∶ y = Av, v ∈ V} (2.16)

5. Inverse linear transformation:

V = A−1 Y ∶= {v ∶ y = Av, y ∈ Y} (2.17)

6. Orthogonal completion1 :

Y = V – ∶= {y ∶ ⟨v, y⟩ = 0, v ∈ V} (2.18)
1
In the coordinate-free subspace algebra a product of vectors v and y is usually denoted as ⟨v, y⟩.
Note that this refers to v T y in the linear matrix algebra.

13
 2. Review

Invariant subspaces

Consider a linear transformation matrix A ∈ Rn×n . A subspace V is A-invariant if and

only if AV ⊆ V. In other words, there exists such a matrix X that (Halmos 1958):

AV = V X (2.19)

Note that if X is diagonal, then columns of V are the eigenvectors of A. The invariance
has a physical meaning in the linear systems theory. Consider an autoregressive system
described by the following equation:

x(t + 1) = Ax(t) (2.20)

It holds that if x(t) ∈ V, where V is A-invariant, then x(t + 1) ∈ V. This means that if
the system is initialised with x(0) ∈ V, then the state vector will remain within V.

Reachability and controllability

Reachability and controllability can be defined by means of geometric tools. Consider

an autoregressive system with an exogenous input:

x(t + 1) = Ax(t) + Bu(t) (2.21)

Using the notation B = Im{B}, the term ⟨A∣B⟩ = B + AB + ⋯ + An−1 B is the infimal
A-invariant subspace containing B, i.e. the reachable subspace of (A, B). This means
that the state trajectory of the system (2.21) driven by the input u(t) can be anywhere
within the reachable subspace of (A, B). Furthermore, because ⟨A∣B⟩ is A-invariant,
the state driven by u(t) cannot leave the reachable subspace of (A, B). Note that, the
reachable subspace of (A, B) can be defined as:

⟨A∣B⟩ = Im{R} (2.22)

where:
R = [ B AB ⋯ An−1 B ] (2.23)

is the reachability matrix of the system.

Analogously, observability can be defined using the geometric theory. Consider the
following autoregressive system:

x(t + 1) = Ax(t)
(2.24)
y(t) = Cx(t)

Denote the kernel of the matrix C as K = Ker{C}, i.e. CK = 0. Then the unobservable
subspace of the system (2.24) is defined as the supremal A-invariant subspace contained

14
 2. Review

in K (Massoumnia 1986):

⟨K∣A⟩ = K ∩ A−1 K ∩ ⋯ ∩ KA−n+1 (2.25)

Note that if the state vector x(t) ∈ K then y(t) = Cx(t) = 0. Because ⟨K∣A⟩ ⊆ K, it
holds that y(t) = Cx(t) = 0 for any x(t) ∈ ⟨K∣A⟩. Furthermore, due to the fact that
⟨K∣A⟩ is A-invariant it holds that x(t + 1) ∈ ⟨K∣A⟩ if x(t) ∈ ⟨K∣A⟩, i.e. the state vector
stays within ⟨K∣A⟩. Thus, if the system (2.24) is initialised with x(0) ∈ ⟨K∣A⟩, then the
state vector remains within the unobservable subspace of (A, C) and the output y(t)
remains zero.
Now consider the relation between ⟨K∣A⟩ and the system observability matrix O:

⎡ ⎤
⎢ ⎥
⎢ ⎥
C
⎢ CA ⎥
⎢ ⎥
O=⎢ ⎥
⎢ ⎥
(2.26)
⎢ ⋮ ⎥
⎢ ⎥
⎢ CAn−1 ⎥
⎣ ⎦

The subspace which is an orthogonal completion of ⟨K∣A⟩ is the observable subspace

of system (2.24). An orthogonal completion of an intersection of two subspaces, V and
Z, is defined as, cf. Equation (3.1.10) in (Basile & Marro 2002):

(V ∩ Z)– = V – + Z – (2.27)

Hence:
⟨K∣A⟩– = K– + (A−1 K) + ⋯ + (A−n+1 K)
– –
(2.28)

Note that the orthogonal subspace of K is Im{C T }. An orthogonal completion of

an inverse transformation of a subspace is defined as, see Property 3.1.3 in (Basile &
Marro 2002):
(A−1 V)– = AT V – (2.29)

Therefore, it holds that:

⟨K∣A⟩– = K– + (A−1 K) + ⋯ + (A−n+1 K) = Im{C T } + Im{AT C T } + ⋯+

– –

(2.30)
+ Im{(An−1 )T C T } = Im{[ C T AT C T ⋯ (An−1 )T C T ]}

which is an image of the transposed observability matrix O. Consequently, the unob-

servable subspace of (A,C) is given by:

⟨K∣A⟩ = Ker {O} (2.31)

15
 2. Review

Controlled and conditioned invariants

The concept of controlled invariants was introduced in (Basile & Marro 1969, Wohnam
& Morse 1970), whilst the concept of conditioned invariance was introduced in (Basile
& Marro 1969). Consider a pair (A, B). A subspace V is an (A, B)-controlled invariant
if:
AV ⊆ V + B (2.32)

This means that there exists such a matrix K that the input u(t) = Kx(t) keeps the
state of the system (2.21) within V, i.e. there exists such a K that (A + BK)V ⊆ V. A
dual of the controlled invariant is a conditioned invariant. Consider a pair (A, C). A
subspace S is said to be an (A, C)-conditioned invariant if:

A(S ∩ K) ⊆ S (2.33)

The (A, C)-conditioned invariance means that there exists such a matrix K that (A −
KC)S ⊆ S. For more properties of controlled and conditioned invariants the reader is
referred to (Basile & Marro 2002).

Invariant zeros

Consider the following system:

x(t + 1) = Ax(t) + Bu(t)

(2.34)
y(t) = Cx(t)

Denote the minimal (A, C)-conditioned invariant containing Im{B} as S0 and use the
notation V0 for the maximal (A, B)-controlled invariant contained in ker{C}. Consider
a matrix V1 such that Im{V1 } ∩ V0 = Im{V1 }, Im{V1 } + V0 = V0 , and Im{V1 } ∩ S0 = 0.
Invariant zeros of (A, B, C) are the eigenvalues of the matrix M1 , which fulfils the
following equation, see (Basile & Marro 2010):

⎡ ⎤
⎢ M1 ⎥
[ V1 ⎢
−B ] ⎢ ⎥ = AV1
⎥
⎢ M2 ⎥
(2.35)
⎣ ⎦

Using a similarity transformation Vm1 , the matrix M1 can be decomposed as:

M1 = Vm−11 JVm1 (2.36)

Equation (2.36) can be, in particular, a Jordan normal decomposition. Consequently,

equation (2.35) can be reformulated as:

V1 Vm−11 JVm1 − BM2 = AV1 (2.37)

16
 2. Review

By postmultiplying both sides of (2.37) by Vm−11 the following formula is obtained:

V1 Vm−11 J − BM2 Vm−11 = AV1 Vm−11 (2.38)

Note that if the invariant zeros of (A, B, C) are distinct, the matrix M1 is diagonalisable,
and J is diagonal, equation (2.38) is equivalent to (2.9), where the diagonal elements of
J are the invariant zeros of (A, B, C), whilst columns of V1 Vm−11 and columns of M2 Vm−11
are, respectively, invariant zeros state and input directions. Therefore, if the invariant
zeros of (A, B, C) are distinct, (2.9) is equivalent to (2.35), where D = 0. Nevertheless,
as opposed to (2.9), Equation (2.35) can be used to determine the number of all of the
invariant zeros of the system, including the repeated ones.
In the case when the system contains a feedthrough term, i.e. D ≠ 0 Basile &
Marro (2002) proposed some manipulations to represent the quadruple (A, B, C, D)
with a triple. However, the geometric approach is used in this thesis to analyse/design
fault detection and isolation filters, where the considered transfer functions between
disturbances or faults and the output of the system do not contain any feedthrough
term. Thus, for more details on invariant zeros of systems with a feedthrough term the
reader is referred to (Basile & Marro 2002).

2.3 Block oriented models

Block-oriented model structures consist of static nonlinearities interconnected with lin-
ear dynamic blocks. In the case of a Hammerstein model, see Fig. 2.1(a), a linear
block is preceded by a static nonlinear function, whereas in the case of a Wiener
model, see Fig. 2.1(c), the order of these elements is reversed (Pearson 1995, Pear-
son & Pottmann 2000). In the case of a Hammerstein-Wiener model structure, cf.
Fig. 2.1(b), a linear dynamic block is preceded and followed by static nonlinearities.
A Wiener-Hammerstein model, cf. Fig. 2.1(d), is characterised by two linear dynamic
blocks connected via a nonlinear static function (Crama & Schoukens 2004). Further-
more, block-oriented systems can be cascaded creating more complicated structures.
Dobrowiecki & Schoukens (2002) studied cascaded Wiener-Hammerstein systems as
the one presented in Fig. 2.2.
u(t) ū(t) y(t) u(t) ū(t) ȳ(t) y(t)
N (⋅) G(z) N1 (⋅) G(z) N2 (⋅)

(a) Hammerstein model (b) Hammerstein-Wiener model

u(t) ȳ(t) y(t) u(t) ū(t) ȳ(t) y(t)

G(z) N (⋅) G1 (z) N (⋅) G2 (z)

(c) Wiener model (d) Wiener-Hammerstein model

Figure 2.1: Block oriented models

17
 2. Review

u(t) y(t)
G1 (z) N1 (⋅) H1 (z) ... Gk (z) Nk (⋅) Hk (z)

Figure 2.2: Cascade of k Wiener-Hammerstein systems

In order to capture system nonlinearities more accurately, a feedback block-oriented

model can be used, see, for example, (Pearson & Pottmann 2000). Different variations
of feedback block oriented models (e.g. a sandwich feedback block-oriented model as
in Fig. 2.3(b), have been studied in (Pottmann & Pearson 2006).

N (⋅) G2 (z) N (⋅) G1 (z)

u(t) y(t) u(t) y(t)

G(z) G0 (z)

(a) Feedback block-oriented model (b) Feedback block-oriented model – a sandwich

structure

Figure 2.3: Feedback block oriented models

2.4 Unknown input reconstruction

It is assumed that the model of a linear system is known and its output is measurable,
however the system is affected by noise. The aim of the unknown input reconstruction
process is to estimate the unknown (unmeasurable) input to the system. The most
trivial solution to this problem, a naive inversion, is not applicable when any system
zero lies outside the unit circle, i.e. the system in nonminimum-phase. Also, due to
highpass properties of a naive inversion, it is not preferable when the system is subjected
to noise. In this Section two input reconstruction algorithms are presented, which
are used as benchmarks for the assessment of the schemes developed in Chapter 3.
In Subsection 2.4.1 a minimum variance unbiased (MVU) state and input estimator
based on the Kalman filter, see (Gillijns & De Moor 2007b), is described. The second
method used as a benchmark is the input estimation (INPEST) method presented in
Subsection 2.4.2.

2.4.1 MVU state estimator

A Kalman filter-based MVU state and input estimator for systems with a direct
feedtrough has been developed by Gillijns & De Moor (2007b). The scheme, for com-
pleteness, is presented in Algorithm 2.1. The system, for which the MVU has been

18
 2. Review

designed, is described by (Gillijns & De Moor 2007b):

x(t + 1) = Ax(t) + Gu(t) + ξ(t)

(2.39)
y(t) = Cx(t) + Hu(t) + ζ(t)

where u(t) is a vector of unknown inputs (the number of unknown inputs is lower or
equal to the number of outputs), whilst ξ(t) and ζ(t) are vectors of white, zero-mean,
Gaussian, i.i.d. noise sequences.
It is assumed throughout this thesis that the noise distribution is Gaussian (i.e.
normally distributed) and zero-mean. Whilst in practice it is known that noise is not
necessary Gaussian, it is commonly accepted that a Gaussian assumption is appropriate;
offering an approach which, although may no longer be optimal, would be consistent
and generalisable, and is applicable to a wide range of situations.

Algorithm 2.1 (MVU).

1. Initialisation

x̂(0) = E{x(0)} (2.40a)

P (0) = E{(x̂(0) − x(0)) (x̂(0) − x(0)) }
x T
(2.40b)
R = E{ζ(t)ζ T (t)} (2.40c)
Q̃ = E{ξ(t)ξ T (t)} (2.40d)

2. Estimation of unknown input

R̃(t) = CP x (t∣t − 1)C T + R(t) (2.40e)

M (t) = (H T R̃−1 H)
−1
H T R̃−1 (2.40f)
û(t) = M (t) (y(t) − C x̂(t∣t − 1)) (2.40g)
P u (t) = (H T R̃−1 H)
−1
(2.40h)

3. Measurement update

K(t) = P x (t∣t − 1)C T R̃−1 (2.40i)

x̂(t∣t) = x̂(t∣t − 1) + K(t) (y(t) − C x̂(t∣t − 1) − H û(t)) (2.40j)
P x (t∣t) = P x (t∣t − 1) − K(t) (R̃(t) − HP u (t)H T ) K T (t) (2.40k)
P xu (t) = (P ux (t))T = −K(t)HP u (t) (2.40l)

19
 2. Review

4. Time update

x̂(t + 1∣t) = Ax̂(t∣t) + Gû(t) (2.40m)

⎡ x ⎤⎡ ⎤
⎢ P (t∣t) P xu (t) ⎥ ⎢ AT ⎥
P x (t + 1∣t) = [ A G ] ⎢ ⎢ ux
⎥⎢
⎥⎢
⎥ + Q̃
⎥ (2.40n)
⎢ P (t) P v (t) ⎥ ⎢ GT ⎥
⎣ ⎦⎣ ⎦

The MVU requires the knowledge of the covariance matrices of the noise sequences,
Q̃ and R, which are the tuning parameters for the algorithm.

Remark 2.1. If the system (2.39) is SISO, the MVU is equivalent to a naive system
inversion.

Demonstration. If (2.39) is a SISO system, H and M are scalars and, therefore,

M = H −1 . Incorporating (2.40g) into (2.40j) it follows that:

x̂(t∣t) = x̂(t∣t − 1) (2.41)

Hence, incorporating (2.40g) into (2.40m):

x̂(t + 1∣t) = (A − M GC)x̂(t∣t − 1) + M Gy(t) (2.42)

Consequently, in the SISO case Algorithm 2.1 is equivalent to:

x̂(t + 1∣t) = (A − M GC)x̂(t∣t − 1) + M Gy(t) (2.43a)

û(t) = −M C x̂(t∣t − 1) + M y(t) (2.43b)

The gain of the filter (2.43) is equal to the reciprocal of the gain of the system, see
(Gillijns & De Moor 2007b). Denote any arbitrary zero of the filter (2.43) as zmvu , then
the following formula holds, cf. (2.9):

⎡ ⎤⎡ ⎤ ⎡ ⎤
⎢ zmvu I − A + M GC −M G ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥⎢ χ ⎥ = ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ −M ⎥ ⎢ κ ⎥ ⎢ 0 ⎥
(2.44)
⎣ −M C ⎦⎣ ⎦ ⎣ ⎦

where a column vector χ denotes the zero state direction and a scalar κ refers to the
input zero direction of zmvu (El-Ghezawi et al. 1983). Consequently:

(zmvu I − A)χ = 0 (2.45)

which means that zmvu is an eigenvalue of A, whereas χ is its corresponding eigenvector.

This means that the zeros of the filter (2.43) are equal to the poles of the system (2.39).

20
 2. Review

Analogously, it can be demonstrated that the poles of (2.43) are equivalent to the
zeros the the system (2.39). Thus, the MVU behaves as a naive inversion in the case
when (2.39) is a SISO system.

2.4.2 Input estimation (INPEST) method

The INPEST algorithm has been designed for the input reconstruction of SISO linear
systems (Young & Sumislawska 2012). A schematic diagram of the INPEST method is
presented in Fig. 2.4. The basic idea of the method is to create a control loop, where
the model of the system is controlled in such a way that its output tracks the measured
output of the real system (which acts as a reference for the control loop). The output
of the controller (i.e. the input to the controlled model) renders the unknown input
estimate. The time delay τ is introduced in order to recognise, that the control system
û(t − τ )

y(t) + ŷ(t − τ )
controller model
_

Figure 2.4: Schematic diagram of INPEST method

does not respond instantaneously to the changes in y(t).

The control system utilises the proportional integral plus (PIP) controller design
that exploits a non-minimum state-space (NMSS) model of the system (Young, Behzadi,
Wang & Chotai 1987). The state vector of the controlled model of the system in the
NMSS representation is defined by:

x(t) = [ ŷ(t) ŷ(t − 1) ⋯ ŷ(t − na + 1) û(t) û(t − 1) ⋯ û(t − nb + 1) z(t) ]

T

(2.46)
where z(t) is an integral of the tracking error:

z(k) = z(k − 1) + (y(k) − ŷ(k)) (2.47)

The PIP controller utilises a state variable feedback (SVF), i.e.

û(t) = −gx(t) (2.48)

where g is the controller gain. The INPEST method makes use of the PIP linear
quadratic (PIP-LQ) design, which minimises the cost function defined as:
∞
J = ∑ {xT (t)Qx(t) + rû2 (t)} (2.49)
t=0

21
 2. Review

where:
Q = diag([ qy ⋯ qy qu ⋯ qu qe ]) (2.50)

and r = qu = qy = 1, whilst qe is optimised. In order to find the optimal value of the qe

parameter, denoted q̊e , and the corresponding estimation delay, τ̊ , the following cost
function J is optimised:

{q̊e ,τ̊ } = arg minJ (qe , τ )

N
J (qe , τ ) = ∑ η 2 (t) + ν(∆û(t))2
t=τ +1 (2.51)
η(t) = y(t) − ŷ(t − τ )
∆û(t) = û(t) − û(t − 1)

where ν is a tuning parameter. The reason for penalising the derivative of the recon-
structed input signal is that in the case of noisy measurements of y(t) the algorithm
would tend to amplify the effects of the disturbance if the rate of change on û(t) was
not penalised. Note that the only tuning parameter of the INPEST method is the
reconstructed input derivative weighting ν.

2.4.3 Efficacy measure of compared algorithms

An efficay index used to assess the ability of the examined algorithms to reconstruct
the unmeasurable signal (unknown input) allows one to determine how much of the
original signal can be explained by the reconstruction algorithm. For this purpose the
following variation of a widely used coefficient of determination, denoted R2T , see, for
example, (Ljung 1999, Young 2011), is utilised:

∑t (û(t) − u(t))2
R2T = × 100%
∑t u2 (t)
(2.52)

where u(t) and û(t) refer to, respectively, the unknown input and its estimate. Note
that in an ideal case, i.e. when û(t) = u(t), R2T = 0 and it increases as the discrepancy
between the original and the estimated input increases.

2.5 Introduction to fault detection and diagnosis

In many industrial processes fault detection and diagnosis cannot be spared as it is
crucial to maintain smooth operation and safety. Undiagnosed or improperly handled
faults can lead to serious consequences, starting from a damage to the product on a
production line (financial loss) up to catastrophic events, which can cost lives. There-
fore, measures should be taken and algorithms implemented, which can isolate and then
deal with (accommodate) the faults. Furthermore, automation and availability of high
computational power increases complexity of industrial systems, which become more

22
 2. Review

vulnerable to faults and hence require a complex monitoring.

An interest in the model-based fault diagnosis started in the early 1970’s with
observer based fault detection (Beard 1971, Jones 1973). Initially, the terminology in
the early fault diagnosis literature was not consistent. In 1991 the SAFEPROCESS
(fault detection, supervision, and safety for technical processes) Steering Committee
was established (in 1993 it became the Technical Committee), which discussed that
matter and formed commonly accepted definitions. Some of the definitions can be
found in (Reliability, Availability, and Maintainability Dictionary 1988) and (Isermann
& Balle 1997). The definitions used throughout this thesis are given in Subsection 2.5.1.
Then, the fault detection problem is discussed in Subsection 2.5.2.

2.5.1 Nomenclature
The fault detection and diagnosis terminology used throughout this thesis is given
below (Isermann & Balle 1997):

Fault: A deviation of at least one characteristic property or parameter of the system

from the acceptable/usual/standard conditions.

Residual: An output of the fault detection/isolation/identification filter. In a fault-

free condition the residual is close to zero and it significantly deviates from zero,
when a fault occurs.

Fault detection: A binary decision, whether a fault is present in the system. Due to
the fact that there is always a certain level of noise in the system a need arises
to distinguish, whether the residual deviates from zero due to the noise or due
to presence of a fault. This is achieved by setting a threshold whose violation
indicates the presence of a fault.

Robust fault detection: A fault detection process which is insensitive to unmea-

sured disturbances. A robust fault detection filter is designed in such a way that
the residuals are insensitive to (decoupled from) disturbances, whilst they are
sensitive to faults.

Fault isolation: Determination of the component which deviates from the accept-
able/usual/standard condition.

Fault identification: Determination of the magnitude of the fault.

Fault diagnosis: Includes fault isolation and identification, i.e. determination of the
source of the fault and the fault magnitude.

23
 2. Review

2.5.2 Fault detection

Consider a process, which can be described by a linear discrete-time time-invariant
model:

x(t + 1) = Ax(t) + Bu(t) + F µ(t)

(2.53)
y(t) = Cx(t) + Du(t)

where x(t) ∈ Rn is the system state vector, u(t) ∈ Rp and y(t) ∈ Rm are, respectively,
the system input and output, and µ(t) ∈ Rk is a fault signal. Matrices A, B, C,
D, and F are constant and have appropriate dimensions. The aim of fault detection
is to define the time instances t, when µ(t) ≠ 0. When both the input and output
measurements are available, so-called process-model-based fault detection methods are
used for the purpose of residual generation (Isermann & Balle 1997, Simani, Fantuzzi
& Patton 2002). These, in particular, are:

1. State and output observers

2. PE

3. Identification and parameter estimation

In this thesis two fault detection methods are considered, which are the state observers
and the PE.

State observer (closed-loop fault detection filter)

A schematic illustration of a state observer is presented in Fig. 2.5 and given by the
following set of equations (Patton 1997, Chen & Patton 1999, Simani et al. 2002):

x̂(t + 1) = Ax̂(t) + (B − KD)u(t) + K(y(t) − ŷ(t))

ŷ(t) = C x̂(t) + Du(t) (2.54)
r(t) = Q(y(t) − ŷ(t))

where K is the observer gain matrix, whilst Q is an arbitrary matrix. (The matrix
Q plays an important role in robust fault detection which is described in Section 2.6;
in this section, however, without loss of generality it is assumed that Q is an identity
matrix.) Consider the state estimation error ξ(t) = x̂(t) − x(t). Then the residual is
governed by:

ξ(t + 1) = (A − KC)ξ(t) + F µ(t)

(2.55)
r(t) = QCξ(t)

24
 2. Review

µ(t)
x(t + 1) = Ax(t) + Bu(t) + F µ(t) y(t)
y(t) = Cx(t) + Du(t)
u(t)

B
+ x̂(t + 1) −1 x̂(t)
C
+ ŷ(t) +
_
r(t)
z + Q
+ +

Figure 2.5: State observer-based (closed-loop) fault detection filter

Therefore, in order to detect an occurrence of a fault, the z-transform transfer function

of the residual response to the fault, denoted as Gµr (z) must be non-zero, i.e.

Gµr (z) = QC(Iz − A + KC)−1 F ≠ 0 (2.56)

Parity equations (open-loop fault detection filter)

PE are widely used for the purpose of fault detection and isolation, see, for example,
(Chow & Willsky 1984, Gertler & Singer 1990, Li & Shah 2002). A schematic illus-
tration of PE is presented in Fig. 2.6. As opposed to the observer-based (closed-loop)
fault detection filters, PE have an open-loop structure.
µ(t)
x(t + 1) = Ax(t) + Bu(t) + F µ(t) y(t)
y(t) = Cx(t) + Du(t)
u(t)

delay delay
_
WQ + W
r(t)

Figure 2.6: Open-loop fault detection filter (PE)

Consider a state-space representation of the system (2.2). The stacked vector of the
system output y(t) is defined as:

Y (t) = [ y T (t − s) y T (t − s + 1) ⋯ y T (t) ]
T
(2.57)

25
 2. Review

where the term s denotes the order of the parity space. Analogously, one can construct
a stacked vector of u(t), which is denoted as U (t). Using this notation the system
defined by (2.53) in a fault-free case can be expressed in the form of:

Y (t) = Γx(t − s) + QU (t) (2.58)

where Γ is an extended observability matrix:

⎡ C ⎤
⎢ ⎥
⎢ ⎥
⎢ CA ⎥
⎢ ⎥
Γ=⎢ ⎥ ∈ R(s+1)m×n
⎢ ⋮ ⎥
(2.59)
⎢ ⎥
⎢ ⎥
⎢ CAs ⎥
⎣ ⎦

and Q is the following block Toeplitz matrix:

⎡ 0 ⎤
⎢ ⋯ ⎥
⎢ ⎥
D 0
⎢ CB 0 ⎥
⎢ ⋯ ⎥
⎢ ⎥
D
⎢
Q = ⎢ CAB 0 ⎥
⎥∈R
(s+1)m×(s+1)p
⋯ (2.60)
⎢ ⎥
CB
⎢ ⋮ ⎥
⎢ ⋮ ⋮ ⋱ ⎥
⎢ ⎥
⎢ CA B CA B D ⎥
⎣ ⋯ ⎦
s−1 s−2

For the purpose of elimination of the unknown state vector from (2.58), a matrix
W ∈ Rl×(s+1)m , l ≥ 1, is defined, which belongs to the left nullspace of Γ, i.e.

WΓ = 0 (2.61)

Note that W can always be found by choosing s to be sufficiently large. Therefore, by

premultiplying (2.58) by W the following expression is obtained:

W Y (t) = W QU (t) (2.62)

Consequently, the residual defined as:

r(t) = W Y (t) − W QU (t) (2.63)

deviates from zero if a fault occurs in the system. It is worth noting that, because the
residual response to fault is open-loop, the residual is correlated only with the last s + 1
samples of the fault signal.

2.6 Robust fault detection

Unmodelled dynamics, process noise, parameter variations, or unmeasurable external
stimuli act as disturbances and may affect the fault detection process leading to false

26
 2. Review

alarms. Therefore, it is required to construct such a fault detection filter, which is

sensitive to faults but insensitive to disturbances. Disturbances are often represented
by an extra input to the system, d(t) ∈ Rq , with a known distribution matrix E ∈
Rn×q . Thus, the system (2.53) with disturbances is represented by the following set of
equations:

x(t + 1) = Ax(t) + Bu(t) + Ed(t) + F µ(t)

(2.64)
y(t) = Cx(t) + Du(t)

A complete disturbance decoupling is achieved if the residual is sensitive to a fault,

i.e. r(t) ≠ 0 if µ(t) ≠ 0, but it is insensitive to disturbances, i.e. if µ = 0 then
r(t) = 0 for any d(t) ∈ R. A robust fault detection filter may be either open-loop
(equivalent to PE) or closed-loop (observer-based). In Subsections 2.6.1 and 2.6.3 the
design of robust, observer-based fault detection filters using, respectively, right and left
eigenstructure assignment (Patton & Chen 1991a, Chen & Patton 1999) is presented.
In Subsection 2.6.2 a geometric insight into a robust fault detection filter via right
eigenstructure assignment is given. It has been shown in (Patton & Chen 1991a) that
a special case of a robust fault detection filter via left eigenstructure assignment is
equivalent to the first order PE, which is shown in Subsection 2.6.4.

2.6.1 Robust fault detection via right eigenstructure assignment

In this subsection a design of a robust fault detection filter using right eigenstructure
assignment is presented. Consider the filter (2.54). The sufficient conditions for the
decoupling of the disturbances from the residual are, see (Chen & Patton 1999):
1. All columns of the matrix E are right eigenvectors of (A − KC) corresponding to
any eigenvalues

2. QCE = 0
It is assumed that E is full column rank. Denote the ith column of E as Ei , which
is also a right eigenvector of (A − KC) corresponding to the desired eigenvalue λi . In
order to satisfy the decoupling condition 1, it should hold that:

(λi I − A + KC)Ei = 0 for i = 1, ⋯, q (2.65)

which is equivalent to:

KCEi = (A − λi )Ei for i = 1, ⋯, q (2.66)

The procedure for finding such a gain matrix K, for which the decoupling conditions
are fulfilled, has been proposed by Chen & Patton (1999). Using the notation:

Aλ = [ (A − λ1 I)E1 (A − λ2 I)E2 ⋯ (A − λq I)Eq ] (2.67)

27
 2. Review

equation (2.66) can be reformulated as:

KCE = Aλ (2.68)

The sufficient conditions to assign all columns of E as right eigenvectors of (A − KC)

are, see (Chen & Patton 1999):
⎡ ⎤
⎢ Aλ ⎥
(i) rank(CE) = rank(⎢
⎢
⎥)
⎥
⎢ CE ⎥
⎣ ⎦
(ii) (C ′ , A′ ) is a detectable pair, where:

A′ = A − Aλ (CE) C
(2.69)
C ′ = (I − CE(CE) )C

The matrix K is subsequently calculated as:

K = Aλ (CE) + K ′ (I − CE(CE) ) (2.70)

where K ′ is an arbitrary matrix (Chen & Patton 1999). Note that:

A − KC = A − Aλ (CE) C − K ′ (I − CE(CE) )C = A′ − K ′ C ′ (2.71)

Consequently, the columns of E are the right eigenvectors of (A′ − K ′ C ′ ) = (A − KC)

corresponding to the desired eigenvectors λi , i = 1, 2, ..., q. Hence, one can allocate re-
maining (n−q) eigenvalues by choosing an appropriate gain matrix K ′ and subsequently
compute the gain matrix K using (2.70). Note that the eigenvalues λi , i = 1, 2, ..., q are
the unobservable modes of the pair (C ′ , A′ ), which means that only remaining n − q
eigenvalues can be allocated by K ′ . The proof of this statement is provided in (Chen
& Patton 1999).

2.6.2 Disturbance decoupling in geometric approach

Consider the robust fault detection filter described in Subsection 2.6.1. Each column
of E is a right eigenvector of (A − KC). Therefore, Im{Ei }, for i = 1, ⋯, q is an (A −
KC)-invariant subspace. Furthermore, Im{Ei }, for i = 1, ⋯, q is an infimal (A − KC)-
invariant subspace containing Im{Ei }, i.e. the reachability subspace of (A − KC, Ei ).
This means that the state trajectory driven by the ith disturbance signal, di (t), will
remain within Im{Ei }. Due to the fact that QCE = 0, Im{Ei } ⊆ Ker{QC} and, because
Im{Ei } is (A−KC)-invariant, it belongs to the supremal (A−KC)-invariant contained
in Ker{QC}, i.e. the unobservable subspace of (A−KC, E, QC). Thus, the state vector
driven by the disturbance remains in the unobservable space of the fault detection filter
(A − KC, E, QC) and, hence, the residual is insensitive to disturbances.

28
 2. Review

2.6.3 Robust fault detection via left eigenstructure assignment

A tutorial paper for the left eigenstructure assignment technique has been written
by (Patton & Chen 1991b). Subsequently, these results have been revisited in (Chen
& Patton 1999). Patton & Chen (1992) utilised the left eigenstructure assignment
technique for a jet engine sensor fault detection.
Any transfer function matrix can be written as:

v1 l1T v2 l2T vn lnT

(zI − A + KC)−1 = + +⋯+ (2.72)
z − λ1 z − λ2 z − λn

where vi and liT are, respectively, the right and left eigenvectors of (A − KC) corre-
sponding to the eigenvalue λi , see (Patton & Chen 1991b, Patton & Chen 1992, Chen
& Patton 1999). Denote left and right eigenvector matrices, respectively, as:

⎡ l1T ⎤
⎢ ⎥
⎢ ⎥
⎢ T ⎥
⎢ l2 ⎥
L=⎢ ⎥ V = [ v1 v2 ⋯ vn ]
⎢ ⋮ ⎥
(2.73)
⎢ ⎥
⎢ ⎥
⎢ lnT ⎥
⎣ ⎦

It is known that the left eigenvector liT is orthogonal to the right eigenvector vj if i ≠ j,
cf. (Patton & Chen 1991b, Chen & Patton 1999). Therefore, if the vectors liT and vi
are appropriately scaled:
LV = I (2.74)

and hence:
L = V −1 (2.75)

The transfer function between the disturbance and the residual can be expressed
as:
n QCvi liT E
Grd (z) = ∑ (2.76)
i=1 z − λi
Hence, Grd (z) vanishes if and only if for i = 1, ⋯, n:

QCvi liT E = 0 (2.77)

which implies that:

n
∑ QCvi li E = QCV LE = QCE = 0
T
(2.78)
i=1

Therefore, the first step for the disturbance decoupling is to find the matrix Q, such that
QCE = 0, see (Patton & Chen 1991b, Chen & Patton 1999). Consequently, sufficient
conditions for the disturbance decoupling using the left eigenstructure assignment are:

1. QCE = 0

29
 2. Review

2. All rows of QC are left eigenvectors of (A−KC) corresponding to any eigenvalues

The proof of the above conditions can be found in (Chen & Patton 1999).

Assignability condition

Rows of QC should be the first rq eigenvectors of (A − KC), where rq denotes the

number of column of QC, i.e.

liT (A − KC) = λi liT (2.79)

where liT is the ith row of QC. The above expression can be reformulated as:

liT (A − λi I) = liT KC (2.80)

Consequently:
liT = −liT KC(A − λi I)−1 (2.81)

Note that liT ∈ R1×n and K ∈ Rn×m . Therefore, liT K ∈ R1×m , whilst the matrix C(A −
λi I)−1 ∈ Rm×n . This means that by premultiplying the matrix C(A − λi I)−1 by a row
vector liT K a linear combination of rows of C(A − λi I)−1 is obtained. Consequently,
a solution to (2.79) exists for the desired λi if and only if the vector liT lies in a
row subspace spanned by C(λi I − A)−1 , i.e. li lies in a column subspace spanned by
(λi I −AT )−1 C T , see (Chen & Patton 1999). Therefore, li must be equal to its projection
on the subspace Im{(λi I −AT )−1 C T }. Denote (λi I −AT )−1 C T as P (λi ). The projection
of li onto Im{(λi I − AT )−1 C T } is given by:

li∗ = P (λi )wi∗ for i = 1, ⋯, rq (2.82)

where:
wi∗ = [P (λi )T P (λi )]−1 P (λi )T li for i = 1, ⋯, rq (2.83)

In the case when li∗ = li , the left eigenvector li is assignable. Otherwise a complete dis-
turbance decoupling using the left eigenstructure assignment is not possible. Consider
the following equation:
(li∗ )T (A − KC) = λi (li∗ )T (2.84)

it holds that:
li∗ = −(λi I − AT )−1 C T K T li∗ (2.85)

Therefore, one can note that, cf. (2.82):

wi∗ = −K T li∗ (2.86)

30
 2. Review

The disturbance decoupling conditions require only rq eigenvectors to be specified

(i.e. rows of QC). The remaining n − rq eigenvectors may be selected freely from
the assignable subspace Im{(λi I − AT )C T }, i.e.:

li = −(λi I − AT )−1 C T wi for i = rq + 1, ⋯, n (2.87)

for any arbitrary wi , i = q + 1, ⋯, n. Subsequently, the gain matrix K is calculated via

(see (Chen & Patton 1999)):
K = −[W L−1 ]T (2.88)

where:
L = [ l1∗ ⋯ lr∗q lrq +1 ⋯ ln ] (2.89)

and:
W = [ w1∗ ⋯ wr∗q wrq +1 ⋯ wn ] (2.90)

2.6.4 Design of first order PE using left eigenstructure assignment

Consider a fault detection filter via left eigenstructure assignment, where (Chen &
Patton 1999):
QC(A − KC) = 0 (2.91)

This occurs when eigenvalues of (A − KC) assigned to the columns of QC are equal
zero. Then, the z-form of the residual is, cf. (2.54):

r(z) = (Q − QC(zI − A + KC)−1 K) y(z)−

(2.92)
(QD − QC(zI − A + KC)−1 (B − QD)) u(z)

where u(z), y(z), and r(z) are the z-transform forms of, respectively, u(t), y(t), and
z(t). Note that:

QC(zI − A + KC)−1 = z −1 QC(I + (A − KC)z −1 + (A − KC)2 z −2 +

(2.93)
(A − KC)3 z −3 + ⋯)

Hence, QC(zI −A+KC)−1 = z −1 QC. Therefore, the computational form of the residual
vector r(z) can be rewritten as:

r(z) = (Q − z −1 QCK) y(z) − (QD − z −1 QC(B − QD)) u(z) (2.94)

which is equivalent to a first order PE:

⎡ ⎤ ⎡ ⎤
⎢ y(t) ⎥ ⎢ ⎥
r(t) = [ Q −QCK ] ⎢
⎢
⎥ − [ QD −QC(B − QD) ] ⎢ u(t)
⎥ ⎢
⎥
⎥
⎢ y(t − 1) ⎥ ⎢ u(t − 1) ⎥
(2.95)
⎣ ⎦ ⎣ ⎦

31
 2. Review

The above scheme is also referred as a deadbeat robust fault detection filter (DRFDF),
see (Chen & Patton 1999).

2.7 Fault isolation and identification

Isermann & Balle (1997) listed various statistical methods, as well as neural networks
and fuzzy logic for fault isolation. This thesis, however, deals with deterministic meth-
ods, which generate residuals of the following properties (Gertler & Kunver 1995):

Structured residual set: Each fault yields certain residuals deviate from zero, whereas
other residuals remain zero. This can be interpreted as the fault µi (t) causing the
residual vector to lie in a certain subspace of the residual space, see Fig 2.7(a).
An example of a structured residual set is presented in Table 2.1.

Fixed direction residuals: Presence of the fault µi (t) yields the residual to lie in a
fixed direction, see Fig 2.7(b). Residual directions do not need to be linearly in-
dependent. However, multiple faults cannot be detected unless residual directions
are linearly independent.

Diagonal residual set: A combination of the two above, i.e. the fault µi (t) causes
the residual ri (t) to deviate from zero, whilst the remaining residuals are equal
to zero. A diagonal residual set can be used to isolate multiple faults. Note that
a diagonal residual set can be obtained from a set of linearly independent fixed
direction residuals by a similarity transformation (change of basis).

r3 (t)
r3 (t)

µ2 (t)
µ3 (t)
µ2 (t)
µ3 (t)
µ1 (t)
r2 (t) r2 (t)
µ1 (t)

r1 (t) r1 (t)
(a) Structured residual set (b) Directional residual set

Figure 2.7: Graphical illustration of structured and directional residual set

2.7.1 Fault isolation via diagnostic observers

Diagnostic observers are state observers which are used for diagnostic purposes. The
pioneers of model-based fault isolation filters are Beard (1971) and Jones (1973). Note

32
 2. Review

Table 2.1: Example of a structured residual set. The entry 1 in the ith row (cor-
responding to the fault µi (t)) and the j th column (corresponding to the
residual rj (t)) denotes that µi (t) yields the residual rj (t) deviate from
zero. Note that multiple faults cannot be isolated using this residual set.

r1 (t) r2 (t) r3 (t) r4 (t)

µ1 (t) 1 0 1 0
µ2 (t) 0 1 1 0
µ3 (t) 1 0 1 1
µ4 (t) 0 1 0 1

that the fault isolation schemes developed by them are often referred to as ‘Beard-Jones
fault detection filters’. In this thesis the term ‘fault detection’ refers to the process of
determining a fault occurrence, whereas the schemes proposed by Beard (1971) and
Jones (1973) are, for sake of consistency, referred to as ‘fault isolation filters’.
Consider the system described by equation (2.53). Denote each column of the matrix
F as Fi and each corresponding fault signal as µi (t). Let δi be the smallest non-negative
integer such that CAδi Fi ≠ 0. Then the term fi = Aδi Fi is further referred to as the
fault direction (Massoumnia 1986, Chen & Speyer 2006a). Consider the following fault
isolation filter:

x̂(t + 1) = (A − KC)x̂(t) + (B − KD)u(t) + Ky(t) (2.96a)

r(t) = y(t) − C x̂(t) − Du(t) (2.96b)

The objectives of the filter design are:

The residual lies in the direction Cfi when the fault µi occurs

Eigenvalues of (A−KC) can be arbitrarily specified (with constraint of conjugate

symmetry and no repeated eigenvalues2 )

In order for the filter design to be feasible, the following conditions must be fulfilled
(Chow & Willsky 1984, Massoumnia 1986, Chen & Speyer 2006b):

1. (C, A) is observable pair

2. rank([ Cf1 Cf2 ⋯ Cfk ]) = k

Assumption 1. ensures that all the filter eigenvalues can be arbitrarily specified, whereas
Assumption 2. allows for the faults to be isolated, i.e. yields residuals caused by
different faults lie in different directions.
2
Constraint of no repeated eigenvalues is often imposed for clarity of analysis and derivation pro-
cesses (Chen & Patton 1999, Chen & Speyer 2006a)

33
 2. Review

Different solutions have been proposed to design a fault diagnostic observer. Mas-
soumnia (1986) represented the Beard-Jones fault isolation filter in a geometric do-
main and, furthermore, added a solution for the filter design when the output re-
sponse to fault has invariant zeros. White & Speyer (1986) reformulated the Beard-
Jones fault isolation filter to an eigenstructure assignment problem. Then, Chen &
Speyer (2006b) used the spectral theory to design a Beard-Jones fault isolation filter.
Furthermore, a design of the filter has been presented using eigenstructure assign-
ment (Chen & Speyer 2006a) and linear matrix inequalities (LMI), see e.g. (Chen &
Nagarajaiah 2007).

Design of fault isolation filter

A fault isolation filter design using the right eigenstructure assignment developed by
Chen & Speyer (2006a) is presented here. Eigenvalues of (A − KC) can be arbitrarily
specified and for each column of F ni eigenvalues, denoted λj , are allocated to (A −
(i)

(i)
KC), corresponding to the eigenvectors wj , j = 1, 2, ⋯, ni :

(A − KC)wj = λj wj
(i) (i) (i)
(2.97)

It is demonstrated in (Chen & Speyer 2006a) that the number of assignable eigenvalues,
ni , depends on the rank of the observability matrix of (Ci , Ai ):

⎡ ⎤
⎢ ⎥
⎢ ⎥
Ci
⎢ CA ⎥
⎢ ⎥
Oi = ⎢ ⎥
i i
⎢ ⎥
(2.98)
⎢ ⋮ ⎥
⎢ ⎥
⎢ Ci Ain−1 ⎥
⎣ ⎦

where Ai = A − Afi (Cfi ) C and Ci = (I − Cfi (Cfi ) ) C and:

ni = n − rank(Oi ) (2.99)

It is also shown that the unobservable subspace of (Ai , Ci ), i.e. ker{Oi }, is

Im{[ Fi AFi ⋯ Aδi Fi ]} ⊕ Vi , where Vi is the subspace spanned by the invariant
zero state directions of (A, Fi , C). Therefore, ni = δi + dim{Vi } + 1. It is pointed out
(i)
that Cfi and Cwj are colinear and, for convenience, it is assumed that (Chen &
Speyer 2006a):
(i)
Cfi = Cwj (2.100)

Incorporating (2.100) into (2.97) the following relation is obtained:

KCfi = (A − λj )wj
(i) (i)
(2.101)

34
 2. Review

The eigenvectors of (A − KC) are defined as:

(i) (i)
wj = Θi βj (2.102)

where Im{Θi } = ker{Oi } and βj

(i)
is a coefficient vector (Chen & Speyer 2006a).
(i)
From (2.100) if follows that the last element of βj is unity, hence (2.102) is refor-

⎡ (i) ⎤
mulated as:
⎢ β̄j ⎥
= [ Θ̄i fi ] ⎢
⎢
⎥ = Θ̄i β̄ (i) + fi
⎥
(i)
⎢ 1 ⎥
wj (2.103)
⎣ ⎦
j

where:
Im{Θ̄i } = Im{[ Fi AFi ⋯ Aδi −1 Fi ]} ⊕ Vi (2.104)

Substituting (2.103) into (2.101):

(A − λj )Θ̄i β̄j = KCfi − (A − λj )fi

(i) (i) (i)
(2.105)

and repeating (2.105) ni times:

⎡ (A − λ(i) )Θ̄ ⎤ ⎡ β̄ (i) ⎤ ⎡ KCf − λ(i) f ⎤

⎢ ⋯ ⎥⎢ 1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ (i) ⎥ ⎢ ⎥
i 0 0 i i
⎢ ⎥ ⎢ β̄ ⎥ ⎢ KCf − λ(i) f ⎥
1 1
⎢ (A − λ2 )Θ̄i ⋯ ⎥⎢ 2 ⎥ ⎢ ⎥
(i)
⎢ ⎥⎢ ⎥=⎢ ⎥(2.106)
0 0 i i
⎢ ⎥⎢ ⋮ ⎥ ⎢ ⎥
2
⎢ ⋮ ⋮ ⋱ ⋮ ⎥⎢ ⎥ ⎢ ⋯ ⎥
⎢ ⎥ ⎢ (i) ⎥ ⎢ ⎥
⎢ ⋯ (A − λni )Θ̄i ⎥ ⎢ β̄ ⎥ ⎢ KCfi − λ(i) ⎥
⎣ ⎦ ⎣ q1 +1 ⎦ ⎣ ⎦
(i)
0 0 n i fi

After subtracting the last row of (2.106) from the others, the following expression is
obtained:
⎡ (A − λ1 )Θ̄i ⎤
−(A − λni )Θ̄i
⎢ ⎥
(i) (i)
⋯
⎢ ⎥
0
⎢ (A − λ2 )Θ̄i ⋯ −(A − λni )Θ̄i ⎥
⎢ ⎥
(i) (i)
⎢ ⎥
0
⎢ ⎥
⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥
⎢ ⋯ (A − λni −1 )Θ̄i − (A − λni )Θ̄i ⎥
⎣ ⎦
(i) (i)
0 0
⎡ β̄1 ⎤ ⎡ ⎤
(2.107)
⎢ ⎥ ⎢ (λ1 − λni )fi ⎥
(i) (i) (i)
⎢ ⎥ ⎢ ⎥
⎢ β̄2 ⎥ ⎢ (λ2 − λni )fi ⎥
⎢ ⎥ ⎢ ⎥
(i) (i) (i)
⎢ ⎥=⎢ ⎥
⎢ ⋮ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⋯ ⎥
⎢ (i) ⎥ ⎢ ⎥
⎢ β̄q1 +1 ⎥ ⎢ (λni −1 − λni )fi ⎥
⎣ ⎦ ⎣ ⎦
(i) (i)

Denote the matrix at the left hand side of (2.107) as Ãi . Then the coefficient vectors
(i)
β̄j , j = 1, 2, ⋯, ni are calculated using a pseudoinverse of Ãi :

⎡ β̄ (i) ⎤ ⎡ (λ(i) − λ(i) )f ⎤

⎢ 1 ⎥ ⎢ ⎥
⎢ (i) ⎥ ⎢ ni i
⎥
⎢ β̄ ⎥ ⎢ (λ(i) − λ(i) )f ⎥
1
⎢ 2 ⎥ ⎢ ⎥
⎢ ⎥ = Ãi ⎢ ⎥
ni i
⎢ ⋯ ⎥ ⎢ ⎥
2
(2.108)
⎢ ⎥ ⎢ ⋯ ⎥
⎢ (i) ⎥ ⎢ (i) ⎥
⎢ β̄ ⎥ ⎢ (λ ⎥
⎣ q1 +1 ⎦ ⎣ ni −1 − λni )fi ⎦
(i)

35
 2. Review

From (2.100) it follows that (Chen & Speyer 2006a):

KCfi = (A − λ1 I)w1 = ⋯ = (A − λ(i)

(i) (i) (i)
ni I)wni (2.109)

Combining the above for i = 1, 2, ⋯, q:

KCf = [ λ(1)
1 w1
(1) (2) (2)
λ1 w 1 ⋯ λ1 w 1 ]
(q) (q)
(2.110)

Consequently, the gain matrix K is calculated as:

K = [ λ(1) ⋯ λ1 w1 ] (Cf ) + K0 (I − (Cf )(Cf ) )

(1) (2) (2) (q) (q)
1 w1 λ1 w 1 (2.111)

where K0 is an arbitrary matrix.

2.7.2 Geometric properties of fault isolation filter

A Beard-Jones fault isolation filter has been derived using a geometric approach in
(Massoumnia 1986). For each fault direction fi it holds that Im{Θi } = Im{[ Fi AFi
⋯ Aδi Fi ]}⊕Vi is an (A−KC)-invariant subspace. Consequently, the state trajectory
driven by the fault µi (t) remains within Im{Θi }. This yields the residual to lie in the
direction Im{CΘi } = Cfi . Due to the fact that Cfi ≠ Cfj , i ≠ j, i.e. different faults
yield different residual directions, faults can be isolated.

2.7.3 Fault identification

In order to identify the magnitude of the fault fault, a fault identification filter is
designed in such a way that the residual approximates the fault signal, i.e. (Ding 2008):

r(t) ≈ µ(t) (2.112)

This is an unmeasurable input reconstruction problem. In the case of multiple faults a

diagonal fault isolation filter can be utilised. It is, however, important that the steady
state gain of the residual response to fault is non-zero (Ding 2008).

2.8 Concluding remarks

In this chapter a background knowledge which is used to develop the algorithms pro-
posed in this thesis has been provided. The reader has been familiarised with the
representation of a dynamic, discrete-time, time-invariant stochastic system in both
polynomial and state-space forms, and an insight into the geometric theory of linear
systems has been given. The problem of an unknown (unmeasurable) input recon-
struction has been introduced and two methods known from the literature have been
presented. Furthermore, the reader has been familiarised with the basics of fault detec-

36
 2. Review

tion and diagnosis. Both closed-loop (observer-based) and open-loop (PE-based) fault
detection/isolation filters have been discussed and appropriate algorithms selected from
the literature have been presented.

37
Chapter 3

Parity equations-based unknown

input reconstruction for linear
stochastic systems

Nomenclature

ai . . . . . . . . . . . . . . autoregressive parameter in polynomial model

A ............... state transition matrix in state-space model
b(t) . . . . . . . . . . . . auxiliary vector
bi . . . . . . . . . . . . . . exogenous parameter in polynomial model
B .............. input matrix of known input in state-space model
ci . . . . . . . . . . . . . . moving average parameter in ARMAX model
C .............. output matrix in a state-space model
D .............. feedforward matrix of known input in state-space model
e(t) . . . . . . . . . . . . noise term
f (⋅) . . . . . . . . . . . . function to be minimised by Lagrange multiplier method
g(⋅) . . . . . . . . . . . . . constraint function in Lagrange multiplier method
G .............. input matrix of unknown input in state-space model
G′ . . . . . . . . . . . . . . auxiliary matrix
Gu (z) . . . . . . . . . . z-domain transfer function between u0 (t) and y(t)
Gv (z) . . . . . . . . . . z-domain transfer function between v(t) and y(t)
G′v (z) . . . . . . . . . . auxiliary z-domain transfer function
H .............. feedforward matrix of unknown input in state-space model
H′ . . . . . . . . . . . . . . auxiliary matrix
k ............... number of rows of the matrix spanning the left nullspace of Γ
K(t) . . . . . . . . . . . gain matrix (used by MVU)
m .............. number of system outputs
M (t) . . . . . . . . . . . auxiliary matrix (used by MVU)
Mv (z) . . . . . . . . . . auxiliary transfer function
n ............... order of system
na . . . . . . . . . . . . . . order of autoregressive polynomial
nb . . . . . . . . . . . . . . order of exogenous polynomial
nc . . . . . . . . . . . . . . order of moving average polynomial

38
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Nv (z) . . . . . . . . . . auxiliary transfer function

p ............... number of known inputs to the system
pi . . . . . . . . . . . . . . element of P
P .............. auxiliary vector
P x (t), P d (t). . . . . submatrices of state and input estimation error covariance matrix
P dx (t), P xd (t) . . . submatrices of state and input estimation error covariance matrix
Q .............. block Toeplitz matrix
Q̃ . . . . . . . . . . . . . . covariance matrix of ξ(t)
R .............. covariance matrix of ζ(t)
R̃ . . . . . . . . . . . . . . covariance matrix (used in MVU)
s ............... parity space order
S ............... auxiliary matrix
T ............... block Toeplitz matrix
T′ .............. auxiliary matrix
u(t) . . . . . . . . . . . . measured input
ũ(t) . . . . . . . . . . . . input measurement noise
ũ∗ (t) . . . . . . . . . . . auxiliary variable
u0 (t) . . . . . . . . . . . noise-free known input
U (t) . . . . . . . . . . . . stacked vector of last s + 1 values of u(t)
U (z) . . . . . . . . . . . u(t) in z-domain
Ũ (t) . . . . . . . . . . . . stacked vector of last s + 1 values of ũ(t)
Ũ (z) . . . . . . . . . . . ũ(t) in z-domain
Ũ ∗ (t) . . . . . . . . . . . stacked vector of last s + 1 values of ũ∗ (t)
Ũ ∗ (z) . . . . . . . . . . ũ∗ (t) in z-domain
U0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of u0 (t)
U0 (z) . . . . . . . . . . u0 (t) in z-domain
v(t) . . . . . . . . . . . . unknown (unmeasurable) input
v̂(t) . . . . . . . . . . . . unknown input estimate
v ′ (t) . . . . . . . . . . . . auxiliary variable
V (t) . . . . . . . . . . . . stacked vector of last s + 1 values of unknown input
V (z) . . . . . . . . . . . v(t) in z-domain
V̂ (z) . . . . . . . . . . . v̂(t) in z-domain
V ′ (t) . . . . . . . . . . . stacked vector of last s + 1 values of v ′ (t)
V ′ (z) . . . . . . . . . . . v ′ (t) in z-domain
w qi , w ξ i . . . . . . . . . auxiliary polynomial parameters
W .............. vector, which belongs to Γ–
W (z) . . . . . . . . . . . polynomial of z-variable defined by appropriate elements of W
WQ (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of W Q
WT (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of W T
WT ′ (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of W T ′
WΞ (z) . . . . . . . . . . polynomial of z-variable defined by appropriate elements of W Ξ
x(t) . . . . . . . . . . . . state vector instate space model
y(t) . . . . . . . . . . . . measured output
y0 (t) . . . . . . . . . . . noise-free output in output-error case
Y (t) . . . . . . . . . . . . stacked vector of last s + 1 values of measured output
Y (z) . . . . . . . . . . . y(t) in z-domain
Y0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of noise-free output
zi . . . . . . . . . . . . . . system zero
αi , αi′ . . . . . . . . . . . auxiliary parameters
βi . . . . . . . . . . . . . . ith diagonal element of Σ in multiple output OE case

39
3. Parity equations-based unknown input reconstruction for linear stochastic systems

χ ............... auxiliary vector

δ ............... system delay
ǫ(t) . . . . . . . . . . . . auxiliary noise term
ǫ∗ (t) . . . . . . . . . . . auxiliary noise term
γ ............... row vector of Γ–
Γ ............... extended observability matrix
Γ– . . . . . . . . . . . . . . left nullspace of Γ
κ ............... auxiliary scalar
λ ............... Lagrange multiplier
ν ............... input derivative weighting
Π ............... input matrix of noise term in the state-space model
Ω ............... feedforward matrix of noise term in the state-space model
Σ, Σe , Σũ , Σũe . . covariance matrices
τ ............... unknown input estimation lag
Ξ ............... block Toeplitz matrix
ψ ............... auxiliary vector
ξ(t) . . . . . . . . . . . . process noise vector in state space model
ζ(t) . . . . . . . . . . . . output noise vector in state space model

Preliminary reading: Sections 2.2, 2.4, and Subsection 2.5.2.

3.1 Introduction
In the literature the problem of the unknown (unmeasurable) input estimation is solved
either by a system inversion or by a joint state and input estimation. Early contribu-
tions to the inversion of multiple-input multiple-output (MIMO) deterministic systems
have been presented by Dorato (1969) and Sain & Massey (1969), however their ap-
proaches did not ensure stability of the inverted systems. Moylan (1977) provided a
stable inversion algorithm for minimum-phase systems, whilst Antsaklis (1978) devel-
oped a straightforward state feedback-based method, which allows to assign poles of
the inverted system. This latter method is however limited to the systems with stable
zeros.
Over the last decade a geometric approach to an unknown input reconstruction has
gained considerable interest, see e.g. (Edelmayer 2005). Kirtikar, Palanthandalam-
Madapusi, Zattoni & Bernstein (2009) proposed an unknown input reconstruction
scheme for minimum phase systems. An exhaustive solution to an unknown-state
unknown-input reconstruction for both minimum-phase and nonminimum-phase sys-
tems has relatively recently been developed by Marro & Zattoni (2010). Nevertheless,
this approach does not consider the effects of measurement noise.
Another approach to the unknown input estimation for deterministic systems is
based on state observers. The Luenberger state observer, see (Luenberger 1964), has
been extended to the class of systems with both, known and unknown system inputs, see
for example (Hou & Müller 1992, Darouach & Zasadzinski 1997). The work of Fernando
& Trinh (2006) presents a joint input and state observer based on a descriptor approach.

40
3. Parity equations-based unknown input reconstruction for linear stochastic systems

When dealing with stochastic systems Kalman filter-based approaches have gained
an interest, see, for example, (Hsieh 2000, Floquet & Barbot 2006). Gillijns & De Moor
(2007a) combined the state observer proposed by Darouach & Zasadzinski (1997) and
the unknown input estimator of Hsieh (2000) creating a joint state and unknown input
observer, which is optimal in the minimum variance sense. This approach has subse-
quently been extended to the case of a linear system with a direct feedthrough term, see
(Gillijns & De Moor 2007b). Palanthandalam-Madapusi & Bernstein (2007) introduced
concept of a state and input observability, i.e. they provided a scheme, which allows
to determine, if both the unknown input and the state can be derived from the output
measurements. Keller & Sauter (2010) proposed a variable geometric Kalman filter,
where the statistical effect of each unknown input is tested before deriving the state
estimate. In the recent work of Ghahremani & Kamwa (2011) an extended Kalman
filter with unknown inputs has been developed and applied to state estimation of a
synchronous machine in a power system.
In this chapter a novel approach to the unknown input reconstruction for MIMO
discrete-time stochastic systems is presented. The parity equation-based unknown in-
put observer (PE-UIO) utilises a parity equations (PE) concept for the unknown input
reconstruction. The design freedom is used to minimise the effect of stochastic distur-
bances on the unknown input estimate. For this purpose a Lagrange multiplier method
is utilised. The proposed method is suitable for both minimum and nonminimum-
phase systems, which is an important result, because unstable zeros may result from
a discretisation of a continuous-time system. The PE-UIO has been originally devel-
oped for single-input single-output (SISO) output error (OE) systems in (Sumislawska,
Burnham & Larkowski 2010). The algorithm has been subsequently extended to the
errors-in-variables (EIV) framework in (Sumislawska, Larkowski & Burnham 2010b).
The analysis of the PE-UIO in frequency domain has been provided in (Sumislawska,
Larkowski & Burnham 2011a). In (Sumislawska, Larkowski & Burnham 2010a) the
scheme has been extended to a MIMO case and a potential application to a steel
rolling mill has been described. In this chapter the PE-UIO is extended to a coloured
process noise case. A generalised form of the algorithm is provided, where the output is
subjected to coloured noise (accounting for measurement and process noise), whilst the
input is affected by white measurement noise. An extension of the PE-UIO algorithm
for the cases when systems zero is close or equal to unity is also provided.
This chapter is organised as follows: in Section 3.2 the problem of the unknown
input reconstruction is stated. Subsequently, in Section 3.3 the PE-UIO is presented.
Then, in Section 3.5, the limitation of the scheme in the case when the system has
zeros close or equal to unity is discussed and an extension, which tackles this problem,
is provided. The proposed algorithms are demonstrated on tutorial examples in Sec-
tion 3.7. Finally, in Section 3.8, the efficacy of the proposed methods is compared with
two existing methods, namely, the minimum variance unbiased (MVU) joint state and

41
3. Parity equations-based unknown input reconstruction for linear stochastic systems

e(t)
v(t)
linear y(t)
u0 (t) system

ũ(t) u(t)

Figure 3.1: Schematic view of a linear system

input estimator, see (Gillijns & De Moor 2007b), and the input estimation (INPEST)
method of Young & Sumislawska (2012).

3.2 Linear system representation

Assume that a linear dynamic discrete-time time-invariant multiple-input multiple-
output (MIMO) system with p known inputs, m outputs and a single unknown input
is represented by an nth order state-space equation of the following form:

x(t + 1) = Ax(t) + Bu0 (t) + Gv(t) + Πe(t)

y(t) = Cx(t) + Du0 (t) + Hv(t) + Ωe(t) (3.1)
u(t) = u0 (t) + ũ(t)

where A ∈ Rn×n , B ∈ Rn×p , C ∈ Rm×n , D ∈ Rm×p , G ∈ Rn×1 , H ∈ Rm×1 , Π ∈ Rn×m ,

Ω ∈ Rm×m . The terms u0 (t) and y(t) refer to, respectively, the known input and
output vectors, whereas v(t) denotes the scalar unknown (unmeasurable) input. The
term e(t) is a column vector of m zero-mean, white Gaussian, i.i.d. noise sequences.
The term ũ(t) is a vector of white, Gaussian, zero-mean, i.i.d. noise sequences, which
is uncorrelated with e(t). The aim of the proposed approach is to reconstruct the
unknown input v(t), minimising at the same time the influence of the disturbances e(t)
and ũ(t) on the estimate. A schematic picture of the considered system is presented in
Fig. 3.1.
Equation (3.1) is a generalised representation of a linear stochastic system and can
be simplified in more specific cases, see Subsection 2.2.3, which, for completeness, are
given below.

ARMAX: The state-space system matrices A, B, C, D, Π, and Ω for the ARMAX

model, cf. (2.5), in the observer canonical form are given by equation (2.6). The
matrices G and H are built by replacing bi in, respectively, matrices B and D
with exogenous matrix parameters related to the unknown input v(t). Note, the
ARMAX model assumes that the input u(t) is known exactly (there is no noise
present on the input variable), hence ũ(t) = 0.

42
3. Parity equations-based unknown input reconstruction for linear stochastic systems

ARX: An ARX model is obtained from the ARMAX model by setting ci , i = 1, ⋯, nc ,

to zero.

OE: An OE case can be modelled by the system representation (3.1), where matrices
A, B, C, D and also G and H are all given as in the ARMAX case. The matrix
Π is null and Ω is diagonal. Also there is no noise present on the input variable,
hence ũ(t) = 0. The PE-UIO algorithm for a SISO OE case has been developed
in (Sumislawska, Burnham & Larkowski 2010).

EIV framework: The EIV framework, see, for example, (Söderström 2007), can be
represented by (3.1), where ũ(t) ≠ 0, Π = 0, and Ω is diagonal. The PE-UIO algo-
rithm for a SISO case in the EIV framework has been presented in (Sumislawska,
Larkowski & Burnham 2010b, Sumislawska et al. 2011a).

3.3 Design of unknown input reconstructor

In this section the PE-UIO algorithm is derived. Firstly, for completeness, the PE
for the state-space model (3.1) are described in Subsection 3.3.1. This is followed by
a development of a new unknown input observer based on PE in Subsections 3.3.2
and 3.3.3. Finally, in Subsection 3.3.4, simplified PE-UIO design algorithms for special
cases of a stochastic linear system, such as SISO OE and MIMO OE, are presented.

3.3.1 Parity equations

The approach presented in this chapter utilises the PE to design an unknown input
reconstructor. Recall the stacked vector of the system output y(t), cf. (2.58):

Y (t) = [ y T (t − s) y T (t − s + 1) ⋯ y T (t) ]
T
(3.2)

where the term s denotes the order of the parity space. Analogously, one can construct
stacked vectors of v(t), u(t), u0 (t), ũ(t) and e(t) which are denoted, respectively, as
V (t), U (t), U0 (t), Ũ (t) and E(t). Using this notation the system defined by (3.1) can
be expressed in the form of:

Y (t) = Γx(t − s) + QU0 (t) + T V (t) + ΞE(t) (3.3)

where Γ is an extended observability matrix, cf. (2.59), and Q is given by (2.60).

Analogously, one can build the matrix T ∈ R(s+1)m×(s+1) by replacing B and D in Q
by, respectively, G and H, and the matrix Ξ ∈ R(s+1)m×(s+1)m is obtained by replacing
B and D in Q by, respectively Π and Ω. The term W ∈ R1×(s+1)m is considered to be a
row vector, which belongs to the left nullspace of Γ, cf. (2.61). Consequently, (3.3) can

43
3. Parity equations-based unknown input reconstruction for linear stochastic systems

be reformulated as, cf. (2.62):

W Y (t) = W QU (t) − W QŨ (t) + W T V (t) + W ΞE(t) (3.4)

By rearranging the measured (known) variables to the right-hand side of (3.4) and the
unknowns to the left-hand side, the following parity relation is obtained, cf. (Li &
Shah 2002):
W T V (t) + W ΞE(t) − W QŨ (t) = W Y (t) − W QU (t) (3.5)

In the next subsection the PE are used in order to derive the PE-UIO.

3.3.2 Unknown input estimation

Denote the matrix spanning the left nullspace of Γ as Γ– . Consequently, the row vector
W is a linear combination of rows of Γ– . In the disturbance-free case, i.e. when
U (t) = U0 (t) and E(t) = 0, the following equation holds, cf. (3.5):

Γ– T V (t) = b(t) (3.6)

where b(t) is a column vector given by:

b(t) = Γ– Y (t) − Γ– QU (t) (3.7)

Selection of a sufficiently large s would lead (3.6) to be a set of equations with an explicit
solution or an overdetermined set of equations. Nevertheless, in practice, precision of
the solution to (3.6) can still be seriously affected by noise. The algorithm proposed here
provides an on-line approximation of the unknown input, simultaneously minimising
unwanted effects of noise.
It is proposed to calculate the value of the unknown input as:

v̂(t − τ ) = W Y (t) − W QU (t) (3.8)

where τ is an estimation lag (estimation delay) and it accounts for the fact that the un-
known input may not be reconstructed instantenously. Therefore, at the time instance
t the estimate of v(t − τ ) is obtained. The estimation delay τ is defined further in this
section. In the noise-free case, v̂(t − τ ) is simply:

v̂(t − τ ) = W T V (t) (3.9)

Therefore, based on the assumption that the unknown input is varying relatively slowly
(see Subsection 3.5.1), its estimate can be calculated as a linear combination of the
sequence v(t − s), v(t − s + 1), ⋯, v(t), i.e.

v̂(t − τ ) = α0 v(t) + α1 v(t − 1) + ⋯ + αs v(t − s) (3.10)

44
3. Parity equations-based unknown input reconstruction for linear stochastic systems

where the α parameters are dependent on the choice of the vector W , such that:

W T = [ αs αs−1 ⋯ α0 ]
T
(3.11)

One can note that (3.10) represents a moving average finite impulse response filter with
the gain being given by the sum of the α parameters, i.e. the sum of elements of the
vector W T . Thus, it is suggested that W should be selected in such a way, that the
sum of elements of the vector W T is equal unity. Furthermore, it is anticipated that
the choice of the order of the parity space s, as well as the vector W , both influence
the estimation lag τ in the estimate of the unknown input (due to the moving average
filtering property of the unknown input estimator). The estimation lag is defined as the
centre of gravity of the moving average filter rounded to the nearest natural number
and is calculated via:
∑ αi i
τ = round ( ) (3.12)
∑ αi
In the following subsection an algorithm for the selection of the optimal vector W
is derived based on the Lagrange multiplier method.

3.3.3 Selection of optimal W

In the case of noisy input and output measurements, equation (3.9) becomes:

v̂(t − τ ) = W T V (t) + W ΞE(t) − W QŨ (t) (3.13)

which can be expanded to give:

v̂(t − τ ) = α0 v(t) + α1 v(t − 1) + ⋯ + αs v(t − s) + wξs+1 e(t) + wξs e(t − 1) + ⋯+

(3.14)
+ wξ1 e(t − s) − wqs+1 ũ(t) − wqs ũ(t − 1) − ⋯ − wq1 ũ(t − s)

where the vector coefficients wξi and wqi , i = 1, 2, ⋯, s + 1, are constructed from the
appropriate elements of the vectors W Ξ and W Q, respectively. (In the case when
p = m = 1, i.e. u(t) and y(t) are scalars, wξi and wqi refer to the ith elements of vectors
W Ξ and W Q, respectively.) Note, that in (3.14) the estimate of the unknown input
is affected by two coloured noise sequences. However, by a careful choice of W , the
degrading effect of these disturbances can be minimised.
Furthermore, the influence of measurement noise on the unknown input estimate
can be reduced by minimising the variance of the term W ΞE(t) − W QŨ (t), i.e.:

E{(W ΞE(t) − W QŨ (t))(W ΞE(t) − W QŨ (t))T } =

(3.15)
= W ΞΣe ΞT W T + W QΣũ QT W T − W ΞΣTũe QT W T − W QΣũe ΞT W T

where Σũ = E{Ũ (t)Ũ T (t)}, Σe = E{E(t)E T (t)}, and Σũe = E{Ũ (t)E T (t)} = 0. Conse-

45
3. Parity equations-based unknown input reconstruction for linear stochastic systems

quently, the vector W should be selected to minimise the cost function f (W ):

f (W ) = W ΞΣe ΞT W T + W QΣũ QT W T (3.16)

subject to the following constraints:

1. Sum of elements of W T is equal to 1.

2. W Γ = 0.

Note, that the condition 1 is sufficient to ensure unity gain, because E{e(t)} = 0 and
E{ũ(t)} = 0.
The cost function (3.16) can be minimised by making use of the Lagrange multiplier
method, see, for example, (Bertsekas 1982). Denote the rows of Γ– by γ1 , γ2 , ..., γk ,
where:
Γ– = [ γ1T ⋯ γkT ]
T
γ2T (3.17)

The vector W is a linear combination of rows of Γ– , which ensures that the constraint
2 is satisfied, i.e.
k
W = ∑ pi γ i (3.18)
i=1

Hence, the cost function (3.16) can be reformulated as a function of the parameter
vector P = [ p1 p2 ⋯ pk ] :
T

k ⎛k ⎞
f (P ) = (∑ pi γi ) Σ ∑ pj γjT
⎝j=1 ⎠
(3.19)
i=1

where
Σ = ΞΣe ΞT + QΣũ QT (3.20)

The cost function f (P ) is required to be minimised subject to the constraint:

g(P ) = sumrow (W T ) − 1 = 0 (3.21)

where the operator sumrow (A) denotes a column vector whose elements are sums of
the appropriate rows of an arbitrary matrix A. (In the case of a row vector q, the term
sumrow (q) is simply a scalar being a sum of elements of the vector q, whilst, if q is a
column vector, sumrow (q) = q.)
The solution to the Lagrange minimisation problem is given by, see (Bertsekas 1982):

∇f (P ) = λ∇g(P ) (3.22)

46
3. Parity equations-based unknown input reconstruction for linear stochastic systems

The cost function (3.19) can be expanded as:

f (P ) = p21 γ1 Σγ1T + p1 p2 γ1 Σγ2T + ⋯ + p1 pk γ1 ΣγkT +

p1 p2 γ2 Σγ1T + p22 γ2 Σγ2T + ⋯ + p2 pk γ2 ΣγkT +
(3.23)
⋮
p1 pk γk Σγ1T + p2 pk γk Σγ2T + ⋯ + p2k γk ΣγkT

Hence, the partial derivative of f (P ) with respect to the ith element of the vector
P (denoted as pi ) is given by:

∂f (P )
=p1 γi Σγ1T + p2 γi Σγ2T + ⋯ + pi γi ΣγiT + ⋯ + pk γi ΣγkT +
∂pi (3.24)
p1 γ1 ΣγiT + p2 γ2 ΣγiT + ⋯ + pi γi ΣγiT + ⋯ + pk γk ΣγiT

Consequently, the gradient of f (P ) can be written as:

⎡ ⎤ ⎡ γ Σγ T γ Σγ T ⋯ γ1 ΣγkT ⎤
⎢ ⎥ ⎢ 1 1 ⎥
∂f (P )
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ γ Σγ T γ Σγ T T ⎥
∂p1 1 2
⎢ ⎥ ⎢ 2 1 ⋯ γ2 Σγk ⎥
∂f (P )
⎢ ⎥=⎢ ⎥P+
2
⎢ ⎥ ⎢ ⎥
∂p2 2

⎢ ⋮ ⎥ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥ ⎢ T ⎥
⎢ ⎥ ⎢ γk Σγ1T γk Σγ2T ⋯ γk Σγk ⎥
⎣ ⎦ ⎣ ⎦
∂f (P )
∂pk
⎡ γ Σγ T γ Σγ T ⋯ γk Σγ1T ⎤
(3.25)
⎢ 1 1 ⎥
⎢ ⎥
⎢ γ Σγ T γ Σγ T T ⎥
2 1
⎢ 1 2 ⋯ γk Σγ2 ⎥
⎢ ⎥ P = (∇f (P ))T
2
⎢ ⎥
2

⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ T ⎥
⎢ γ1 Σγ T γ2 Σγ T ⋯ γk Σγk ⎥
⎣ k k ⎦

Thus, recalling that Σ is symmetric, expression (3.25) can be reformulated as:

(∇f (P ))T = (Γ– Σ (Γ– ) + (Γ– Σ (Γ– ) ) ) P

T T T
(3.26)

The constraint function g(P ) is:

k
g(P ) = sumrow (W T ) − 1 = ∑ sumrow (pi γi T ) − 1 (3.27)
i=1

Hence the partial derivative of g(P ) with respect to pi is calculated via:

∂g(P )
= sumrow (γi T ) (3.28)
∂pi

Thus, the gradient of g(P ) can be reformulated as:

(∇g(P ))T = sumrow (Γ– T ) (3.29)

47
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Using the notation:

S = Γ– Σ(Γ– )T + (Γ– Σ (Γ– )T )
T
(3.30)

and
ψ = sumrow (Γ– T ) (3.31)

the solution to the Lagrange optimisation problem (3.22) can be rewritten as:

SP = λψ (3.32)

Hence, the optimal parameter vector P is given by:

P = λS −1 ψ (3.33)

The constraint function g(P ) = 0 can be rewritten as:

PTψ − 1 = 0 (3.34)

Incorporating (3.33) into (3.34) yields:

λ (S −1 ψ) ψ − 1 = 0
T
(3.35)

Finally, the Lagrange multiplier is given by:

λ = ((S −1 ψ) ψ)
T −1
(3.36)

Consequently, the algorithm for calculating the optimal vector W and estimation of
the unknown input is summarised as follows:

Algorithm 3.1 (PE-UIO).

1. Select the order of the parity space s ≥ n and build matrices Γ, Q, T , and Ξ.

2. Obtain Γ– .

3. Compute Σ as:

Σ = ΞΣe ΞT + QΣũ QT (3.37a)

4. Calculate the column vector S via:

S = Γ– Σ(Γ– )T + (Γ– Σ (Γ– )T )

T
(3.37b)

48
3. Parity equations-based unknown input reconstruction for linear stochastic systems

5. Compute the matrix ψ by making use of:

ψ = sumrow (Γ– T ) (3.37c)

6. Obtain the Lagrange multiplier λ:

λ = ((S −1 ψ) ψ)
T −1
(3.37d)

7. Calculate the parameter vector P by:

P = λS −1 ψ (3.37e)

8. Compute the vector W using:

W = P T Γ– (3.37f)

9. Calculate the estimation lag as:

∑ αi i
τ = round ( ) (3.37g)
∑ αi

where:
W T = [ αs αs−1 ⋯ α0 ]
T

10. Obtain the estimate of v(t − τ ) via:

v̂(t − τ ) = W Y (t) − W QU (t) (3.37h)

It should be noted that, due to the fact that the arg min f (P ) needs to be found,
cf. (3.19), the function f (P ) can be scaled by an arbitrary number. Therefore, the
covariance matrices of ũ(t) and e(t) do not require to be known explicitly. It is sufficient
to know only the ratio between the variances of the noise sequences and scale Σũ and
Σe accordingly.

3.3.4 Design of PE-UIO for OE systems

In the case when the system input measurements are noise-free and there is no process
noise, whilst the output vector is affected by a white, Gaussian, zero-mean, i.i.d. noise
sequences (OE case), the procedure of finding the optimal vector W can be simplified,
which is presented in this subsection.

49
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Single-output OE

Consider a single-output OE system described by (3.1). Without loss of generality it

is assumed that Ω = 1 and e(t) is a white, zero-mean, Gaussian sequence with the
variance of var(e(t)). Since there is no process noise and Ξ is an identity matrix, the
term Σ, cf. (3.20), is given by:

Σ = Σe = var(e(t))I (3.38)

Since the objective is to find the minimum of the cost function f (P ), it can be scaled
by any arbitrary number. Therefore, for sake of simplicity, the term var(e(t)) can be
omitted. Consequently, the cost function f (P ) becomes, cf. (3.19):

k ⎛k ⎞
f (P ) = (∑ pi γi ) ∑ pj γjT
⎝j=1 ⎠
(3.39)
i=1

One can select Γ– such that its rows are orthonormal, i.e. Γ– (Γ– ) = I. Therefore, the
T

cost function can be reformulated as:

k
f (P ) = ∑ p2i (3.40)
i=1

This gives a partial derivative if f (P ) equal to:

∂f (P )
= 2pi (3.41)
∂pi

Consequently:
(∇f (P ))T = 2P (3.42)

Incorporating (3.29) and (3.42) into (3.22) the solution to the Lagrange optimisation
problem is calculated as:
P = λsumrow (Γ– ) (3.43)

The constraint equation (3.27) can be reformulated as:

P T sumrow (Γ– ) = 1 (3.44)

Incorporating (3.43) into (3.44) yields:

λ (sumrow (Γ– )) sumrow (Γ– ) = 1

T
(3.45)

Consequently, the Lagrange multiplier is calculated as:

λ = [(sumrow (Γ– )) sumrow (Γ– )]

T −1
(3.46)

50
3. Parity equations-based unknown input reconstruction for linear stochastic systems

A simplified version of Algorithm 3.1 for single-output OE systems is given below:

Algorithm 3.2 (PE-UIO for single-output OE).

1. Select the order of the parity space s ≥ n and build matrices Γ, Q, T , and Ξ.

2. Obtain Γ– such that Γ– (Γ– ) = I.

T

3. Obtain the Lagrange multiplier λ using (3.46).

4. Calculate the parameter vector P by (3.43).

5. Compute the vector W using (3.18).

6. Compute τ using (3.12).

7. Obtain the estimate of v(t − τ ) via equation (3.8).

Multiple output OE

Consider the system (3.1), where m > 1, ũ(t) = 0, Π = 0 and Ω is diagonal. In such
a case Ξ and consequently Σ, cf. (3.20), are diagonal matrices. Assume that Γ– is
selected, such that its rows are orthonormal vectors, i.e. Γ– (Γ– ) = I. Then the cost
T

function f (P ) can be simplified to, cf. (3.40):

k
f (P ) = ∑ βi p2i (3.47)
i=1

where βi denotes the ith element of the diagonal of Σ. Therefore, the partial derivative
of f (P ) is calculated as:
∂f (P )
= 2βi pi (3.48)
∂pi
Consequently, the gradient of f (P ) is:

(∇f (P ))T = 2ΣP (3.49)

Therefore, the Lagrange optimisation problem can be reformulated as, cf. (3.43):

ΣP = λsumrow (Γ– ) (3.50)

Incorporating (3.50) into (3.44) yields:

λ (sumrow (Γ– )) Σ−1 sumrow (Γ– ) = 1

T
(3.51)

51
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Consequently, the Lagrange multiplier is calculated as:

λ = ((sumrow (Γ– )) Σ−1 sumrow (Γ– ))

T −1
(3.52)

A simplified version of Algorithm 3.1 for multiple-output OE systems is given below:

Algorithm 3.3 (PE-UIO for multiple-output OE).

1. Select the order of the parity space s ≥ n and build matrices Γ, Q, T , and Ξ.

2. Obtain Γ– such that Γ– (Γ– ) = I.

T

3. Obtain the Lagrange multiplier λ using (3.52).

4. Calculate the parameter vector P by:

P = λΣ−1 sumrow (Γ– ) (3.53)

5. Compute the vector W using (3.18).

6. Compute the estimation lag τ using (3.12).

7. Obtain the estimate of v(t − τ ) via equation (3.8).

3.4 Analysis in frequency domain

The two relationships between each system input (both known and unknown) and the
output can be described by discrete-time transfer functions of the z-variable of the
following form, cf. (3.1):

Gu (z) = C(zI − A)−1 B + D

(3.54)
Gv (z) = C(zI − A)−1 G + H

Denote y(t), u0 (t), and v(t) in the z-domain, respectively, as Y (z), U0 (z), and V (z).
Consequently, equation (3.4) in the noise-free case can be reformulated as the following
relation:
W (z)Y (z) = WQ (z)U0 (z) + WT (z)V (z) (3.55)

where terms W (z), WQ (z) and WT (z) are appropriate polynomial vectors of the z-
variable with parameters defined by vectors W , W Q, and W T , respectively. Therefore,
in the noise-free case, the relationship between the unknown input and its estimate in

52
3. Parity equations-based unknown input reconstruction for linear stochastic systems

the z-domain is given by, cf. (3.10) and (3.11):

V̂ (z) = W (z)Y (z) − WQ (z)U0 (z) = WT (z)V (z) (3.56)

In the case when noise is present in the system, equation (3.56) becomes:

V̂ (z) = WT (z)V (z) + WΞ (z)E(z) − WQ (z)Ũ (z) (3.57)

where WΞ (z) and E(z) refer to, respectively, the appropriate polynomial vector of the
z-variable with parameters defined by the vector W Ξ and the variable e(t) in z-domain,
whilst Ũ (z) denotes the z-domain representation of ũ(t).
In the case when p = m = 1 the transfer functions corresponding to u0 (t) and v(t)
are given, respectively, by:

WQ (z)
Gu (z) =
W (z)
WT (z)
(3.58)
Gv (z) =
W (z)

where Gu (z) defines the relationship between U0 (z) and the output, whereas Gv (z)
describes the relationship between V (z) and Y (z), cf. (3.54). In the case when s = n,
the left nullspace of Γ is a row vector Γ– = W (it is assumed here that the system (3.1)
is observable) and the degree of the polynomial W (z) is equal to the order of the
system. Hence, one can deduce from (3.58) that the roots of the polynomial W (z) are
eigenvalues of the matrix A (i.e. poles of both Gv (z) and Gu (z)). Denote the set of
poles and zeros of Gv (z) by Pv and Zv , respectively. Analogously, refer to Pu and Zu
as, respectively, poles and zeros of Gu (z). Then, it is true that the roots of W (z) are
Pv ∪ Pu , the roots of WQ (z) are defined by the set Zu ∪ (Pv /Pu ), whilst roots of WT (z)
are Zv ∪ (Pu /Pv ).
If the order of the parity space is higher than that of the system, i.e. s > n, then
the set of equations (3.58) must still be fulfilled. This means, that W (z), WQ (z) and
WT (z) have common s − n roots (a zero-pole cancellation occurs, hence both
WQ (z)
W (z)
WT (z)
and W (z) remain unaltered). The choice of those additional s − n zeros influences
the properties of the noise filtration of the filter (3.13). Hence, the problem of finding
the optimal vector W can be reformulated as a filter zeros assignment problem. The
unknown input reconstruction is possible when the bandwidth of the unknown input is
narrower than that of WT (z), whilst the ability of the PE-UIO to filter ũ(t) and e(t)
depends on the frequency response of both, i.e. WQ (z) and W (z).

53
3. Parity equations-based unknown input reconstruction for linear stochastic systems

3.5 Two stage PE-UIO

The previous section explains, why the PE-UIO cannot be used when the Gv (z) con-
tains a derivative term, i.e. a zero equal to unity. In such a case the polynomial WT (z)
also contains the derivative term and its steady state gain is zero. Therefore, use of
the standard PE-UIO for the purpose of the unknown input estimation is infeasible.
Furthermore, if Gv (z) contains a zero close to unity, the step response of WT (z) is
characterised by a large overshoot (characteristic for systems whose zeros lie close to
unity), hence the unknown input estimate becomes seriously affected. The overshoot
of WT (z) can be minimised by a significant increase of the order of the parity space,
however this results in a reduction of the bandwidth of the filter. Therefore, a modifica-
tion of the PE-UIO is needed, and it is provided in the next subsection. Note that this
problem will occur also for multiple input systems as long as the system has a single
output. Thus, during the derivation of the modified PE-UIO filter in Subsection 3.5.1
it is assumed that the single output system may have an arbitrary number of measured
inputs, i.e. m = 1 and p ∈ N. The algorithm developed in the following section is appli-
cable to systems, whose ‘problematic’ zero lies on the real axis and is lower or equal to
unity. This result is extended in Section 3.6 to the cases with multiple zeros which lie
on or within the unit circle and are relatively close to unity.

3.5.1 Two stage filter design

Consider a single output system, whose transfer function between the unknown input
and the output, denoted as Gv (z), contains a zero, denoted as z0 , which is close or
equal to unity. Such a transfer function can be represented by:

z − z0
Gv (z) = G′v (z) (3.59)
z

Therefore, the input-output relationship in a z-domain can be represented as:

z − z0
Y (z) = Gv (z)V (z) + Gu (z)U0 (z) = G′v (z) V (z) + Gu (z)U0 (z) =
z (3.60)
= G′v (z)V ′ (z) + Gu (z)U0 (z)

where V ′ (z) = z−z0

z V (z) is the z-domain representation of the variable v ′ (t), whose
relation with the unknown input is defined as:

v ′ (t) = v(t) − z0 v(t − 1) (3.61)

The transfer function G′v (z) can be represented by:

G′v (z) = C(zI − A)−1 G′ + H ′ (3.62)

54
3. Parity equations-based unknown input reconstruction for linear stochastic systems

where H ′ and G′ are the appropriately modified matrices H and G, respectively. The
matrix T ′ is calculated by replacing G and H in T by, respectively, G′ and H ′ . Subse-
quently, (3.5) can be reformulated as:

W T ′ V ′ (t) + W ΞE(t) − W QŨ (t) = W Y (t) − W QU (t) (3.63)

Analogously to the algorithm described in Section 3.3, it is proposed to estimate the

variable v ′ (t − τ ) as:
v̂ ′ (t − τ ) = W Y (t) − W QU (t) (3.64)

which in the noise-free case is equal to:

v̂ ′ (t − τ ) = W T ′ V ′ (t) (3.65)

Subsequently, the unknown input estimate can be calculated via, cf. (3.61):

v̂(t) = z0 v̂(t − 1) + v̂ ′ (t) (3.66)

Note that this scheme is applicable only to systems with ∣z0 ∣ ≤ 1. Otherwise, (3.66)
becomes unstable.
In the noisy case the term v̂ ′ (t) is given by:

v̂ ′ (t − τ ) = W T ′ V ′ (t) + ǫ(t) (3.67)

where ǫ(t) accounts for the disturbance introduced by e(t) and ũ(t), i.e.:

ǫ(t) = W ΞE(t) − W QŨ (t) (3.68)

Hence, it follows from equations (3.66) and (3.67), that the estimate of v(t − τ ) is
affected by the error term ǫ∗ (t), whose relation to ǫ(t) is given by:

ǫ∗ (t) = ǫ(t) + z0 ǫ∗ (t − 1) (3.69)

For convenience, the following notation is introduced:

ũ∗ (t) = ũ(t) + z0 ũ∗ (t − 1)

(3.70)
e∗ (t) = e(t) + z0 e∗ (t − 1)

Thus, the term ǫ∗ (t) is given by:

ǫ∗ (t) = W ΞE ∗ (t) − W QŨ ∗ (t) (3.71)

where terms E ∗ (t) and Ũ ∗ (t) are built from the current and previous values of e∗ (t)
and ũ∗ (t), respectively, cf. (3.2). It is required to minimise the variance of the term

55
3. Parity equations-based unknown input reconstruction for linear stochastic systems

ǫ∗ (t), which is given by:

var(ǫ∗ (t)) = E{(W ΞE ∗ (t) − W QŨ ∗ (t))(W ΞE ∗ (t) − W QŨ ∗ (t))T } =

(3.72)
= W ΞΣe∗ ΞT W T + W QΣũ∗ QT W T − W Ξ (Σũ∗ e∗ )T QT W T − W QΣũ∗ e∗ ΞT W T

where Σũ∗ = E{Ũ ∗ (t)(Ũ ∗ (t))T }, Σe∗ = E{E ∗ (t)E ∗ T (t)} and Σũ∗ e∗ = E{Ũ ∗ (t)E ∗ T (t)} =
0. Hence, the function to be minimised is given by:

f (W ) = W ΞΣe∗ ΞT W T + W QΣũ∗ QT W T (3.73)

In order to calculate (3.73), first, the terms Σe∗ and Σũ∗ need to be obtained. The
signal e∗ (t) can be described by a function of its previous values, cf. (3.70):

e∗ (t) = e(t) + z0 e(t − 1) + z02 e(t − 2) + ⋯ (3.74)

Therefore, by recalling that e(t) is assumed white, the expected value of e∗ (t)e∗ (t − i)
is calculated as:

E{e∗ (t)e∗ (t − i)} = E{z0i e2 (t − i) + z0i+2 e2 (t − i − 1) + z0i+4 e2 (t − i − 2) + ⋯}

(3.75)
= E{e2 (t)}z0i (1 + z02 + z04 + ⋯)

which is a sum of a geometric series and in the case when ∣z0 ∣ < 1 it can be simplified
to:

z0i
E{e∗ (t)e∗ (t − i)} = E{e2 (t)} (3.76)
1 − z02

Analogously, by recalling that ũ(t) is assumed white, the expected value of ũ∗ (t)ũ∗ (t−1)
can be derived as:

z0i
E{ũ∗ (t)(ũ∗ (t − i))T } = E{ũ(t)ũT (t)} (3.77)
1 − z02

(Note that e(t) is a scalar, whilst ũ(t) is, in general, a vector.) In the case when ∣z∣ = 1
the sum of the geometric series (3.75) is infinite. Therefore, to cope with such a case
it is proposed to replace z0 in (3.76) and (3.77) by a value smaller than unity in order
to indicate that e∗ (t) and ũ∗ (t) are not white.
The matrices Σe∗ and Σũ∗ are built by filling their entries by the appropriate values
of, respectively, E{e∗ (t)e∗ (t − i)} and E{ũ∗ (t)ũ∗ (t − i)}. For convenience, a new term
is introduced, cf. (3.20):
Σ∗ = ΞΣe∗ ΞT + QΣũ∗ QT (3.78)

56
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Hence, the cost function (3.73) becomes:

s−n+1 ⎛s−n+1 ⎞
f (P ) = ( ∑ pi γiT ) Σ∗ ∑ pj γ j
⎝ j=1 ⎠
(3.79)
i=1

which is required to be minimised subject to the constraint:

g(P ) = sumrow (W T ′ ) − 1 = 0 (3.80)

The solution to this constrained optimisation problem is solved analogously to the one in
Section 3.3. Therefore, the algorithm for calculating the unknown input is summarised
as follows:

Algorithm 3.4 (Two stage PE-UIO).

1. Select the order of the parity space s ≥ n and build matrices Γ, Q, T ′ , and Ξ.

2. Obtain Γ– .

3. Calculate Σe∗ and Σũ∗ using (3.76) and (3.77), respectively.

4. Compute Σ∗ using (3.78).

5. Calculate the matrix S by:

S = Γ– Σ∗ (Γ– )T + (Γ– Σ∗ (Γ– )T )T (3.81a)

6. Calculate the column vector ψ via:

ψ = sumrow (Γ– T ′ ) (3.81b)

7. Obtain the Lagrange multiplier λ via:

λ = ((S −1 ψ) ψ)
T −1
(3.81c)

8. Calculate the parameter vector P by:

P = λS −1 ψ (3.81d)

9. Compute the vector W , cf. (3.18), as:

W = P T Γ– (3.81e)

57
3. Parity equations-based unknown input reconstruction for linear stochastic systems

10. Calculate the estimation lag as:

∑ αi′ i
τ = round ( ) , for i = 0, ⋯, s (3.81f)
∑ αi′

where αi parameters are defined by the equation:

W T ′ = [ αs′ αs−1 ⋯ α0′ ]

T
′ (3.81g)

11. Obtain v̂ ′ (t − τ ) via:

v̂ ′ (t − τ ) = W T ′ V ′ (t) (3.81h)

12. Obtain the estimate of v(t − τ ) as:

v̂(t − τ ) = z0 v̂(t − τ − 1) + v̂ ′ (t − τ ) (3.81i)

3.5.2 Analysis in frequency domain

The variable v ′ (t − τ ) in the z-domain is given by, cf. (3.64):

V ′ (z) = WT ′ (z)V ′ (z) − WQ (z)Ũ (z) + WΞ (z)E(z) (3.82)

where the coefficients of the polynomial WT ′ (z) are appropriate elements of the vector
W T ′ . Consequently, the unknown input estimate in the z-domain is, see (3.57):

V̂ (z) = WT ′ (z)V (z) − WQ (z)Ũ (z) + WΞ (z)E(z)

z z
(3.83)
z − z0 z − z0

It can be deduced that the use of the two stage PE-UIO is advisable if z0 is a positive
real number lower or equal unity. Firstly, if the single stage PE-UIO is used in such
a case, the presence of z0 in WT (z) will cause an overshoot in the step response of
z
the input estimation filter, which may me undesirable. Secondly, the factor z−z0 , for
0 ≤ z0 ≤ 1, reduces the bandwidth of the noise affecting the input estimate, cf. (3.83).
On the other hand the use of the single stage PE-UIO may be preferred over its two
stage version if z0 > 0 is relatively close to zero, and the phase lead caused by the
presence of z0 is desirable, e.g. in an on-line application, when the fast response of the
filter is required. It is not recommended to use the two stage PE-UIO in noisy systems
z
when z0 is lower than zero, due to highpass properties of z−z0 , which would cause an
amplification of noise effect on the unknown input estimate.

58
3. Parity equations-based unknown input reconstruction for linear stochastic systems

3.6 Generalised two stage PE-UIO

The two stage PE-UIO presented in Section 3.5 is used to cope with a single ‘problem-
atic’ zero. It is worth exploring a generalised form of the-two stage PE-UIO, which can
eliminate more than one ‘inconvenient’ zero. Therefore, the transfer function Gv (z)
can be formulated as, cf. (3.59):

Gv (z) = Mv (z)Nv (z) (3.84)

where Nv (z) is in a form of:

∑ki=1 (z − zi )
Nv (z) = (3.85)
zk
and k denotes the number of zeros which need to be eliminated from the unknown input
reconstruction filter. In general it is assumed that zeros z1 , ⋯, zk are complex (with the
constraint of conjugate symmetry). Similarly, as in Section 3.5, zeros z1 , ⋯, zk must be
stable or marginally stable. Consider a variable v ′ (t), cf.(3.61), which in z-domain is
given by:
V ′ (z) = Nv (z)V (z) (3.86)

Consequently:
Gv (z)V (z) = Mv (z)V ′ (z) (3.87)

The transfer function Mv (z) is defined as:

Mv (z) = C(zI − A)−1 G′ + H ′ (3.88)

where H ′ and G′ are the appropriately modified matrices H and G, respectively. The
matrix T ′ is calculated by replacing G and H in T by, respectively, G′ and H ′ . Subse-
quently, in the first stage of the algorithm, the term V ′ (z) is estimated as:

V̂ ′ (z) = W (z)Y (z) − WQ (z)U (z) (3.89)

The unknown input is calculated in the second stage of the algorithm as:

V (z) = Nv−1 (z)V ′ (z) (3.90)

where Nv−1 (z) is defined as:

zk
Nv−1 (z) = (3.91)
∑ki=1 (z − zi )
In the case when noise is present in the system, the variable V ′ (z) is given by, cf.
(3.64), (3.67) and (3.68):

V̂ ′ (z) = WT′ (z)V ′ (z) + WΞ (z)E(z) − WQ (z)Ũ (z) (3.92)

59
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Therefore, the unknown input estimate is, cf. (3.90):

V̂ (z) = WT′ (z)V (z) + W (z)Ξ Nv−1 (z)E(z) − WQ (z)Nv−1 (z)Ũ (z) (3.93)

The following notation is used:

E ∗ (z) = Nv−1 (z)E(z) (3.94)

The term E ∗ (z) is denoted in time domain as e∗ (t) and, subsequently, the stacked
vector of e∗ (t) is E ∗ (t). Analogous notation is used for Nv−1 Ũ (z), i.e. Ũ ∗ (z), ũ∗ (t),
and Ũ ∗ (t).
Analogously to (3.72), the variance of the following term must be minimised:

E{(W ΞE ∗ (t) − W QŨ ∗ (t))(W ΞE ∗ (t) − W QŨ ∗ (t))T } =

(3.95)
= W ΞΣe∗ ΞT W T + W QΣũ∗ QT W T − W Ξ (Σũ∗ e∗ )T QT W T − W QΣũ∗ e∗ ΞT W T

where Σũ∗ = E{Ũ ∗ (t)(Ũ ∗ (t))T }, Σe∗ = E{E ∗ (t)E ∗ T (t)} and Σũ∗ e∗ = E{Ũ ∗ (t)E ∗ T (t)} =
0. Hence, the function to be minimised is given by:

f (W ) = W Σ∗ W T (3.96)

where:
Σ∗ = ΞΣe∗ ΞT + QΣũ∗ QT (3.97)

The covariance matrices Σe∗ and Σũ∗ depend on zi , i = 1, ⋯, k and variances of e(t) and
ũ(t) and should be calculated for each case individually. Finally, the generalised two
stage PE-UIO is summarised as follows:

Algorithm 3.5 (Generalised two stage PE-UIO).

1. Select zeros, z1 , ⋯, zk , which need to be eliminated from the PE.

2. Calculate Nv (z) using (3.85).

3. Select the order of the parity space s ≥ n and build matrices Γ, Q, T ′ , and Ξ.

4. Obtain Γ– .

5. Calculate Σe∗ and Σũ∗ .

6. Compute Σ∗ using (3.97).

7. Calculate the matrix S by:

S = Γ– Σ∗ (Γ– )T + (Γ– Σ∗ (Γ– )T )T (3.98a)

60
3. Parity equations-based unknown input reconstruction for linear stochastic systems

8. Calculate the column vector ψ via:

ψ = sumrow (Γ– T ′ ) (3.98b)

9. Obtain the Lagrange multiplier λ as:

λ = ((S −1 ψ) ψ)
T −1
(3.98c)

10. Calculate the parameter vector P by:

P = λS −1 ψ (3.98d)

11. Compute the vector W as:

W = P T Γ– (3.98e)

12. Calculate the estimation lag using (3.81f).

13. Obtain v̂ ′ (t − τ ) as:

v ′ (t − τ ) = W Y (t) − W QU (t) (3.98f)

14. Obtain the estimate of the unknown input using (3.90).

3.7 Tutorial examples

In this section the design of the proposed approaches is demonstrated on numerical
examples. In Subsection 3.7.1 the design of the PE-UIO is demonstrated on two ex-
amples, namely, an OE case as well as an ARMAX system in the EIV framework.
The influence of the choice of the tuning parameter s on the frequency response of
the filter is also presented. In Subsection 3.7.2 the design of the two stage PE-UIO is
demonstrated and compared with the standard PE-UIO.
Although the examples presented here are described by feedtrough models, it should
be noted that the causality of physical systems assumes that there is a delay on the
system input, denoted as δ, which results in the transfer function defined by:

bn z −δ + bn−1 z −1−δ + ⋯ + b0 z −n−δ

G(z −1 ) = (3.99)
an + an−1 z −1 + ⋯ + a0 z −n

where z −1 is a backwards shift operator, i.e. z −1 y(t) = y(t − 1), G(z −1 ) denotes the
transfer function between any of the system input (either v(t) or u(t)) and the output,

61
3. Parity equations-based unknown input reconstruction for linear stochastic systems

ai and bi , i = 1, ⋯, n, are transfer function polynomial coefficients, and n refers to the

order of the system. Without loss of generality, the delay term δ can be omitted and
therefore, the transfer function G(z −1 ) can be represented by:

bn + bn−1 z −1 + ⋯ + b0 z −n
G(z −1 ) = (3.100)
an + an−1 z −1 + ⋯ + a0 z −n

which corresponds to the following z-domain transfer fuction:

bn z n + bn−1 z n−1 + ⋯ + b0
G(z) = (3.101)
an zn + an−1 z n−1 + ⋯ + a0

3.7.1 PE-UIO
This subsection presents a step-by-step design of the PE-UIO algorithm for two different
cases, namely, an OE and an ARMAX case in the EIV framework. The importance of
the tuning parameter s on the frequency response of the filter is also explained.

Example 3.1. Design of the PE-UIO in the OE case

Consider a linear system, whose transfer functions, Gu (z) and Gv (z), cf. (3.54) are
given by:

z + 0.01
Gu (z) =
(z − 0.9)(z − 0.85)
(z + 1.95)(z − 0.2)
(3.102)
Gv (z) =
(z − 0.9)(z − 0.85)

It can be represented by equation (3.1), whose matrices are given by:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ 1.750 1 ⎥ ⎢ 1.00 ⎥ ⎢ 3.500 ⎥
⎢
A=⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎥ B=⎢ ⎥ G=⎢ ⎥
⎢ −0.765 0 ⎥ ⎢ 0.01 ⎥ ⎢ −1.155 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ (3.103)
C =[ 1 0 ] D=0 H =1

It is assumed that the output of the system is affected by a white, zero-mean, Gaussian
noise sequence of the variance var(e(t)) = 1. Since the considered case is single-output
OE, Algorithm 3.3 is used for the unknown input reconstruction.
The parity space order is chosen to be s = 4, hence the extended observability matrix

⎡ ⎤
is:
⎢ ⎥
⎢ ⎥
1.0000 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
1.7500 1
Γ=⎢
⎢ 2.2975 1.7500 ⎥
⎥
⎢ ⎥
(3.104)
⎢ 2.6819 2.2975 ⎥
⎢ ⎥
⎢ ⎥
⎢ 2.9357 2.6819 ⎥
⎣ ⎦

62
3. Parity equations-based unknown input reconstruction for linear stochastic systems

The left nullspace of Γ is calculated as:

⎡ −0.3578 0.1454 0.8094 −0.2794 −0.3431 ⎤

⎢ ⎥
– ⎢⎢ ⎥
Γ = ⎢ −0.3118 0.4635 −0.2389 0.5947 −0.5264 ⎥
⎥
⎢ ⎥
(3.105)
⎢ −0.2720 0.6999 −0.2723 −0.4955 0.3412 ⎥
⎣ ⎦

Note that Γ– is orthonormal, i.e. Γ– (Γ– )T = I. The matrix Q is given by:

⎡ 0 0 ⎤
⎢ ⎥
⎢ ⎥
0 0 0
⎢ 0 0 ⎥
⎢ ⎥
⎢ ⎥
1 0 0
Q=⎢
⎢ 0 0 ⎥
⎥
⎢ ⎥
1.7600 1 0 (3.106)
⎢ 0 0 ⎥
⎢ 2.3150 1.7600 1 ⎥
⎢ ⎥
⎢ 2.7049 2.3150 1.7600 1 0 ⎥
⎣ ⎦

whereas the matrix T is:

⎡ ⎤
⎢ ⎥
⎢ ⎥
1.0000 0 0 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
3.5000 1.0000 0 0 0
T =⎢
⎢
⎥
⎥
⎢ ⎥
4.9700 3.5000 1.0000 0 0 (3.107)
⎢ ⎥
⎢ 6.0200 4.9700 3.5000 1.0000 0 ⎥
⎢ ⎥
⎢ 4.9700 3.5000 1.0000 ⎥
⎣ 6.7330 6.0200 ⎦

Using (3.46) the Lagrange multiplier λ is calculated to be 0.0403. Subsequently, the

parameter vector P is obtained using (3.43):

⎡ −0.1608 ⎤
⎢ ⎥
⎢ ⎥
P =⎢ ⎥
⎢ −0.1199 ⎥
⎢ ⎥
(3.108)
⎢ 0.0083 ⎥
⎣ ⎦

The vector W is calculated as, cf. (3.18):

W = [ 0.0927 −0.0732 −0.1038 −0.0305 0.1211] (3.109)

Consequently, the vector W T is :

W = [ −0.0480 0.1463 0.3856 0.3929 0.1232 ] (3.110)

which corresponds to τ = 1, cf. (3.12). Therefore, v̂(t − 1) is calculated via:

v̂(t − 1) = 0.1211y(t) − 0.0305y(t − 1) − 0.1038y(t − 2) − 0.0732y(t − 3)+

(3.111)
0.0927y(t) + 0.1211u(t) + 0.1827u(t − 1) + 0.1229u(t − 2) + 0.0012u(t − 3)

63
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Example 3.2. Design of the PE-UIO for an ARMAX system in the EIV
framework
Consider an ARMAX system, whose A, B, C, D, G, and H matrices are the same as
in Example 3.1. The moving average coloured noise parameters are c0 = 1, c1 = 0.3,
and c2 = 0.1. It is assumed that the input measurement is affected by a white, zero-
mean, Gaussian noise ũ(t) of the variance equal to the variance of e(t). Thus the noise
distribution matrix is given by:

⎡ ⎤
⎢ 2.050 ⎥
⎢
Π=⎢ ⎥
⎥
⎢ −0.665 ⎥
(3.112)
⎣ ⎦

whereas Ω = 1. The parity space order is selected as s = 4. Matrices Q, T , Γ, and Γ–

are the same as in Example 3.1. Then, one can compute the matrix Ξ as:

⎡ 0 ⎤
⎢ ⎥
⎢ ⎥
1 0 0 0
⎢ 0 ⎥
⎢ ⎥
⎢ ⎥
2.0500 1 0 0
Ξ=⎢
⎢ 0 ⎥
⎥
⎢ ⎥
2.9225 2.0500 1 0 (3.113)
⎢ 0 ⎥
⎢ 3.5461 2.9225 2.0500 1 ⎥
⎢ ⎥
⎢ 3.5461 2.9225 2.05 1 ⎥
⎣ 3.9700 ⎦

Note that the variances of ũ(t) and e(t) are unknown, however for the purpose of
finding the optimal filter parameters only the ratio between those variances is needed,
which is equal to one. Substituting unity for the variances of both ũ(t) and e(t), the
term Σ is calculated using (3.37a):

⎡ 3.9700 ⎤
⎢ ⎥
⎢ ⎥
1.0000 2.0500 2.9225 3.5461
⎢ 14.3895 ⎥
⎢ ⎥
⎢ ⎥
2.0500 6.2025 9.8011 12.5071
Σ=⎢
⎢ 28.8699 ⎥
⎥
⎢ ⎥
2.9225 9.8011 17.8411 24.2391 (3.114)
⎢ 44.5789 ⎥
⎢ 3.5461 12.5071 24.2391 35.7753 ⎥
⎢ ⎥
⎢ 58.8525 ⎥
⎣ 3.9700 14.3895 28.8699 44.5789 ⎦

whereas S is equal to, cf. (3.37b):

⎡ 5.6976 3.5814 −0.2238 ⎤

⎢ ⎥
⎢ ⎥
S=⎢
⎢ −0.3369 ⎥
⎥
⎢ ⎥
3.5814 2.8100 (3.115)
⎢ −0.2238 −0.3369 ⎥
⎣ 1.0128 ⎦

Consequently, the vector ψ is given by, cf. (3.37c):

⎡ −3.9894 ⎤
⎢ ⎥
⎢ ⎥
ψ=⎢
⎢ −2.9748 ⎥
⎥
⎢ ⎥
(3.116)
⎢ 0.2060 ⎥
⎣ ⎦

64
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Using (3.37d), the Lagrange multiplier is calculated as λ = 3.1271 and the parameter
vector P is computed from (3.37e) as:

⎡ −0.0477 ⎤
⎢ ⎥
⎢ ⎥
⎢
P = ⎢ −0.2749 ⎥
⎥
⎢ ⎥
(3.117)
⎢ −0.0383 ⎥
⎣ ⎦

Finally, the vector W is obtained using (3.37f):

W = [ 0.1132 −0.1612 0.0375 −0.1312 0.1480 ] (3.118)

The estimation lag is equal to 1. Therefore, the unknown input estimate v̂(t − 1) is
calculated via:

v̂(t − 1) = 0.1480y(t) − 0.1312y(t − 1) + 0.0375y(t − 2) − 0.1612y(t − 3)+

0.1132y(t − 4) + 0.1480u(t − 1) + 0.1293u(t − 2)+ (3.119)
0.1492u(t − 3) + 0.0015u(t − 4)

Importance of the tuning parameter s

The order of the parity space s is a tuning parameter of the PE-UIO algorithm. It is
anticipated that an increase of s will lead to a reduction of the impact of disturbances
on the unknown input estimate. At the same time it is expected that an increase
of the order of parity space will yield a reduction of the filter bandwidth, which will
result in the input reconstruction filter being sluggish. This phenomenon can be seen
in Fig. 3.2, where the frequency responses of the polynomial filters WT (z), WQ (z), and
W (z), cf. (3.55), for three different cases of s are compared. This effect is also visible
in Fig. 3.3, where the reconstructed input signals are compared for different values of
parity space orders. The system from Example 3.1 is considered in this experiment.
Whilst for s = 4 the unknown input estimate is noisy (i.e. the noise filtering is rather
poor in this case), for s = 15 the filter does not reproduce high frequency oscillations
of the input. The PE-UIO with s = 7 seems to be the optimal setting for the given
example.

3.7.2 Two stage PE-UIO

In this subsection a design of the two stage PE-UIO is presented on a numerical example.
Furthermore, using different scenarios, the efficacy of the algorithm is compared with
that of the standard PE-UIO.

65
3. Parity equations-based unknown input reconstruction for linear stochastic systems

W (z)
0
s=4
−20 s=7
s=15
−40

−60 −1 0
10 10
WQ (z)
0
Magnitude [dB]

−20

−40

−60 −1 0
10 10
WT (z)
0

−20

−40

−60 −1 0
10 10
Frequency [rad/s]

Figure 3.2: Frequency responses of W (z), WQ (z) and WT (z) for different values of
the parity space order s

15 true
Unknown input estimate

s=4
s=7
10
s=15
5

−5

−10

−15
4750 4800 4850 4900 4950
Time [samples]

Figure 3.3: Comparison of unknown input estimates for different values of s

Example 3.3. Design of the two stage PE-UIO for an OE model

Consider a linear system, whose transfer functions, Gu (z) and Gv (z) are given by:

z − 0.1
Gu (z) =
(z − 0.9)(z − 0.8)
(z + 1.2)(z − 0.95)
(3.120)
Gv (z) =
(z − 0.9)(z − 0.8)
66
3. Parity equations-based unknown input reconstruction for linear stochastic systems

It can be represented by the state-space model (3.1), whose matrices are:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ 1.70 1 ⎥ ⎢ 1 ⎥ ⎢ 1.95 ⎥
⎢
A=⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎥ B=⎢ ⎥ G=⎢ ⎥
⎢ −0.72 0 ⎥ ⎢ −0.1 ⎥ ⎢ −1.86 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ (3.121)
C =[ 1 0 ] D=0 H =1

It is assumed that the output of the system is subjected to white, zero-mean, Gaussian
measurement noise (OE case) of the variance var(e(t)) = 2.7. After elimination of the
zero at 0.95, the corresponding modified matrices H ′ and G′ are built such that H ′ = 1

⎡ ⎤
and:
⎢ 2.90 ⎥
⎢
G =⎢ ⎥
⎥
′
⎢ −0.72 ⎥
(3.122)
⎣ ⎦
The covariance of e∗ (t) is calculated as, cf. (3.76):

0.95i
E{e∗ (t)e∗ (t − i)} = var(e(t)) (3.123)
1 − 0.952

Consequently:

⎡ 1.0000 0.9500 0.9025 0.8574 0.8145 0.7738 ⎤

⎢ ⎥
⎢ ⎥
⎢ 0.9500 1.0000 0.9500 0.9025 0.8574 0.8145 ⎥
⎢ ⎥
⎢ ⎥
⎢ 0.9025 0.9500 1.0000 0.9500 0.9025 0.8574 ⎥
⎢ ⎥
= 2.7 ⎢ ⎥
1
⎢ ⎥
Σe∗ (3.124)
⎢ 0.8574 0.9025 0.9500 1.0000 0.9500 0.9025 ⎥ −
1 0.952
⎢ ⎥
⎢ 0.8145 0.8574 0.9025 0.9500 1.0000 0.9500 ⎥
⎢ ⎥
⎢ ⎥
⎢ 0.7738 0.8145 0.8574 0.9025 0.9500 1.0000 ⎥
⎣ ⎦

The vector W is calculated to be:

W = [ 0.0658 −0.0576 −0.0417 −0.0292 −0.0196 0.0914 ] (3.125)

which corresponds to τ = 3. Therefore, the unknown input is computed via:

v ′ (t − 3) = 0.0914y(t) − 0.0196y(t − 1) − 0.0292y(t − 2) − 0.0417y(t − 3)−

0.0576y(t − 4) + 0.0658y(t − 5) + 0.0914u(t − 1) + 0.1267u(t − 2)+
(3.126)
0.1223u(t − 3) + 0.0779u(t − 4) − 0.0091u(t − 5)
v(t) = v ′ (t) + 0.95v(t − 1)

Fig. 3.4 compares step responses of WT ′ (t) for the two stage PE-UIO with s = 5
and WT (t) for two cases of the standard PE-UIO (s = 5 and s = 15). It can be noted
that the system zero at 0.95 causes a large overshoot in the case when the standard
PE-UIO is used (for both s = 5 and s = 15). This results in a significant distortion of the
unknown input estimate, which can be seen in Fig. 3.5, where the time-domain result of

67
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Step response of WT (z) and WT ′ (z)

7
standard PE−UIO, s=5
6 standard PE−UIO, s=15
two−stage PE−UIO, s=5
5

0
0 5 10 15
Time [samples]

Figure 3.4: Step responses of WT ′ (t) for the two stage PE-UIO and WT (t) for two
cases of the standard PE-UIO

true
60
standard PE−UIO, s=5
Unknown input estimate

standard PE−UIO, s=15

40 two−stage PE−UIO, s=5

20

−20

1.27 1.275 1.28 1.285 1.29 1.295 1.3 1.305

Time [samples] x 10
4

Figure 3.5: Comparison of the unknown input estimate using the two stage PE-UIO
and the standard PE-UIO, z0 = 0.95. Distortion of the unknown input
estimate caused by the phase lead can be seen in the case of the standard
PE-UIO.

the unknown input reconstruction in a noise-free case is shown. The advantage of the
two stage PE-UIO in this particular case can be also seen in Fig. 3.6, where frequency
responses of W (z), WQ (z), WT (t), and WT ′ (t) are presented.

Example 3.4. Comparison of the standard and the two stage PE-UIO in an
on-line application
In this example a situation is presented, where the phase lead introduced by the PE-
UIO is advantageous. Consider an on-line application, where an estimation delay is
crucial for the system performance, e.g. where the reconstructed input is utilised by
a feedback controller. Bearing in mind that at the time instance t the delayed input

68
3. Parity equations-based unknown input reconstruction for linear stochastic systems

W (z)

−20
standard PE−UIO s=5
−40 standard PE−UIO s=15
two−stage PE−UIO s=5
−60 −1 0
10 10
WQ (z)
20
Magnitude [dB]

0
−20
−40
−60 −1 0
10 10
WT (z) and WT′ (z)
20
0
−20
−40
−60 −1 0
10 10
Frequency [rad/s]

Figure 3.6: Frequency responses of WT ′ (t) for the two stage PE-UIO and WT (t) for
two cases of the standard PE-UIO

estimate v̂(t − τ ) is obtained, the on-line estimation error is defined as:

ǫon−line (t) = v̂(t − τ ) − v(t) (3.127)

(In contrary, in an off-line situation or when the estimation delay is not crucial the
difference between v̂(t − τ ) and v(t − τ ) is taken into consideration.)
The system used in this example is given by the equation:

(z − 0.3)(z + 1.8)
Gv (z) =
(z − 0.8)(z − 0.9)
(3.128)

The zero at z0 = 0.3 causes a phase lead (and consequently an overshoot of the step
response) of WT (z) when the standard PE-UIO is used. However, it is expected that
the zero at 0.3 will reduce the impact of the estimation lag caused by the zero at −1.8.
(Due for the fact that 0.3 lies relatively far from unity, it is not expected to cause
such as damaging distortion in the step response of WT (z) as shown in Example 3.3.)
Consequently, a faster response is anticipated when using the PE-UIO instead of the two
stage PE-UIO with z0 = 0.3, which may be particularly desired in on-line applications.
In Fig. 3.7, for completeness, step responses of WT (z), for the standard PE-UIO,

69
3. Parity equations-based unknown input reconstruction for linear stochastic systems

and WT ′ (z), for the two stage PE-UIO, are compared (the parity space order is in both
cases s = 2).

2
standard PE−UIO
two−stage PE−UIO

Step response of WT (z)

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3
Time [samples]

Figure 3.7: Step responses of WT (z) and WT ′ (z), s = 2, z0 = 0.3

Sample time responses of the unknown input estimates using the two algorithms
are presented in Fig. 3.8. It is assumed that the output of the system is subjected

On-line estimation error

1
0
−1
5300 5305 5310 5315 5320 5325 5330 5335 5340
Unknown input estimate
10
true
9 standard PE−UIO
two−stage PE−UIO
8

2
5300 5305 5310 5315 5320 5325 5330 5335 5340
Time [samples]

Figure 3.8: On-line unknown input estimation using the standard PE-UIO and the
two stage PE-UIO. In both cases order of parity space s = 2, variance of
noise var(e(t)) = 14e-4.

to low level OE noise (var(e(t)) = 14e-4). It can be noted that the input estimate,
when using the standard PE-UIO, yields a smaller estimation delay compared to the

70
3. Parity equations-based unknown input reconstruction for linear stochastic systems

two stage PE-UIO (τ = 0 in the case of the PE-UIO and τ = 1 for the two stage
PE-UIO), which is due to the fact that, when the standard PE-UIO is used, the lead
caused by the zero at 0.3 partially compensates for the lag caused by the zero at −1.8.
Nevertheless, the two stage PE-UIO has superior noise filtering properties (in terms
of the bandwidth of W (z)), what can be observed in Fig. 3.9. The efficacy of the

W (z)
10
standard PE−UIO
0
two−stage PE−UIO
−10
−20
−30
−40 −1 0
10 10
WT (z)
5
0
−5
−10
−15 −1 0
10 10
Frequency [rad/s]

Figure 3.9: Frequency responses of WT (z) and WT ′ (z), s = 2, z0 = 0.3

two algorithms for different levels of noise and different orders of the parity space s
are compared in Table 3.1. It can be noted that for relatively low levels of noise the
standard PE-UIO preforms better in terms of the on-line input estimation error (3.127)
than the two stage PE-UIO, due to the lag compensation. However, as the noise level
increases, the on-line input estimation error variance increases more slowly when the
two stage PE-UIO is utilised, which is due to superior noise filtering properties of the
two stage PE-UIO. As the order of the parity space, i.e. s, is increased, the level of the
OE noise, for which both the standard PE-UIO and the two stage PE-UIO perform the
same in terms of the variance of the on-line input estimation error, is also increased.

3.8 Comparison with other methods

In this section the efficacy of the PE-UIO is compared with two other algorithms,
namely, the minimum variance unbiased (MVU) state and input estimator, see (Gillijns
& De Moor 2007b), and the INPEST (input estimation), see (Young & Sumislawska
2012). In order to assess the efficacy of the considered algorithms, the R2T , cf. Sec-
tion 2.4.3, is used.

Example 3.5. Comparison of efficacy of the PE-UIO and the MVU

In this example the efficacy of the unknown input reconstruction using two methods,

71
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Table 3.1: Comparison the standard PE-UIO and the two stage PE-UIO in terms
of the variance of the on-line unknown input reconstruction error, i.e.
ǫon−line (t), for different levels of noise and different orders of the parity
space. The term ‘% std dev’ refers to the percentage value of the ratio
between the output meausrement noise e(t) and the system output y0 (t)
in terms of the standard deviation.

s var(e(t)) % std dev 1-stage PE-UIO 2-stage PE-UIO

2 14e-4 0.0099 0.0855 0.2994
2 2.8e-1 0.1407 0.4056 0.4082
2 14 0.99 16.1863 5.7649
4 14e-1 0.3145 1.1337 1.5725
4 14 0.9947 1.8377 1.9169
4 14e1 3.1455 8.8554 5.3312
6 14 0.9947 2.5897 3.0861
6 14e1 3.1455 3.8324 3.8192
6 14e2 9.9469 16.1160 11.0143

namely, the PE-UIO the MVU is compared. The following single-input two-output
ARX model is considered:
⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
y(t) = ⎢ ⎥ y(t − 1) + ⎢ 0.76 0 ⎥ y(t − 2)+
−1.75 0
⎢ ⎥ ⎢ ⎥
⎢ −1.75 ⎥ ⎢ 0 0.765 ⎥
⎣ ⎦ ⎣ ⎦
0
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ (3.129)
⎢ 1⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ v(t) + ⎢ 1.3 ⎥ v(t − 1) + ⎢ 2.4 0 ⎥ e(t − 1)
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 1⎥ ⎢ −0.3 ⎥ ⎢ 0 0.7 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦

which corresponds to the system (3.1), whose matrices are given by:

⎡ 1.75 1 0 ⎤ ⎡ 3.05 ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 2.4 0
⎢ 0 ⎥ ⎢ ⎥ ⎢ 0.7 ⎥
⎢ 0 1 ⎥ ⎢ 1.45 ⎥ ⎢ ⎥
A=⎢ ⎥ G=⎢ ⎥ Π=⎢ ⎥
1.75 0
⎢ −0.76 ⎥ ⎢ ⎥ ⎢ 0 ⎥
⎢ 0 0 ⎥ ⎢ −0.76 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 0
⎢ 0 ⎥ ⎢ ⎥ ⎢ 0 ⎥
⎣ 0 0 ⎦ (3.130)
−0.765 ⎣ −0.765 ⎦ ⎣ 0 ⎦
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ 1 0 0 0 ⎥ ⎢ 1 ⎥ ⎢ 0 0 ⎥
⎢
C =⎢ ⎥ H =⎢ ⎥ Ω=⎢ ⎥
⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 1 0 0 ⎥ ⎢ 1 ⎥ ⎢ 0 0 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦

Note, that the model (3.129) corresponds to the model (2.39), whose A, B, G, and H
matrices are as in (3.130), whilst ζ(t) = 0 and:

⎡ ⎤
⎢ ⎥
⎢ ⎥
2.4 0
⎢ 0.7 ⎥
⎢ ⎥
ξ(t) = ⎢ ⎥ e(t)
0
⎢ 0 ⎥
(3.131)
⎢ ⎥
⎢ ⎥
0
⎢ 0 ⎥
⎣ 0 ⎦

72
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Consequently, the noise covariance matrices used by the MVU are R = 0 and:

⎡ 0 0 ⎤
⎢ ⎥
⎢ ⎥
5.76 0
⎢ 0.49 0 0 ⎥
⎢ ⎥
Q̃ = var(e(t)) ⎢ ⎥
0
⎢ 0 0 ⎥
(3.132)
⎢ ⎥
⎢ ⎥
0 0
⎢ 0 0 ⎥
⎣ 0 0 ⎦

The algorithms are compared for tree different levels of var(e(t)), namely, 1, 8, and
0.001, which correspond to, respectively, 4.8 %, 13.6 %, and 0.152 % of noise on each
output by means of the standard deviation. In the experiment the MVU is compared
with the PE-UIO designed with different values of s. A Monte-Carlo simulation with
100 runs is carried-out in order to provide reliable results, which are presented in Ta-
ble 3.2. The MVU ensures the minimum variance of the estimation error resulting from

Table 3.2: Comparison of efficacy of PE-UIO and MVU

var(e(t)) 1 8 0.001
s τ R2T [%] τ R2T [%] τ R2T [%]
2 0 2.2153 0 17.6747 0 0.0091202
3 1 1.7598 1 9.4901 1 0.6559958
4 1 0.9371 1 6.0884 1 0.2019992
5 2 1.5533 2 5.4152 2 1.0017866
6 2 1.5191 2 4.6083 2 1.0776268
7 3 2.4092 3 4.9843 3 2.0409051
MVU 0 1.9986 0 15.7028 0 0.0019627

the disturbances. Therefore, achieving lower R2T than that of MVU and ensuring at the
same time τ = 0 is not feasible, what can be seen in the simulation results. However, the
major advantage of the PE-UIO is the ability to adjust the filter bandwidth by selecting
the tuning parameter s. By choice of an optimal s, the R2T is reduced approximately
2 and 4 times for, respectively, var(e(t)) = 1 and var(e(t)) = 8 compared to the MVU.
The results show that for a low level of noise (var(e(t)) = 0.001) the MVU performs
better than the PE-UIO. This is due to the fact that the PE-UIO provides an estimate
of the unknown input, cf. (3.10) and (3.11).

Example 3.6. Comparison of the INPEST, MVU, standard PE-UIO, and

two stage PE-UIO
In this example four methods are compared, namely, the standard PE-UIO, the two
stage PE-UIO, the INPEST and the MVU. Due to the fact that the INPEST method
has been designed for SISO systems, the considered model has only one unknown input
and no known inputs:
z 2 + 0.55z − 0.38
Gv (z) = (3.133)
z 2 + 0.05z − 0.756

73
3. Parity equations-based unknown input reconstruction for linear stochastic systems

The output of the system is subjected to white, zero-mean, Gaussian noise of unity
variance. Note that in the SISO case, the MVU resembles a naive inversion, cf. Re-
mark 2.1. Results of 100-run Monte-Carlo simulation are presented in Table 3.3, whilst
samples of the estimated input signals are presented in Fig. 3.10. It can be noted

Error
2
0
−2
1.13 1.132 1.134 1.136 1.138 1.14 1.142 1.144 1.146 1.148 1.15 1.152
4
Unknown input estimate x 10

true input
10 MVU
PE−UIO
2 stage PE−UIO
INPEST

−5

−10

1.13 1.132 1.134 1.136 1.138 1.14 1.142 1.144 1.146 1.148 1.15
Time [samples] x 10
4

Figure 3.10: Comparison of unknown input reconstruction efficacy of standard PE-

UIO, two stage PE-UIO, INPEST, and MVU

that the standard PE-UIO, the two stage PE-UIO, and the INPEST provide compa-
rable results, whereas the MVU seems to give inferior results in terms of R2T . This is
due to the relatively high bandwidth of the MVU, which results in the lowest possible
estimation lag (in this case τ = 0). The other examined algorithms can be tuned to
reduce the reconstruction filter bandwidth (by increasing ν in the case of the INPEST
method and s in the case of the standard and the two stage PE-UIO), which also yields
an inherent estimation delay (τ > 0).

Example 3.7. Comparison of the INPEST and the two stage PE-UIO
In this example the efficacy of the two stage PE-UIO and the INPEST is compared in
the case when the output response to unknown input contains a zero close to unity.

74
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Table 3.3: Results of comparison of various input reconstruction methods. Noise vari-
ance var(e(t)) = 1

Method R2T [%] τ

PE-UIO (s = 4) 1.1538 1
2-stage PE-UIO (s = 5) 1.1395 2
INPEST 1.3557 1
MVU 5.5980 0

The considered system is described by the following transfer function:

0.01759z 2 + 0.05856z − 0.07367

Gv (z) = (3.134)
z 2 − 1.868z + 0.8706

The system has two zeros at: −4.3023 and 0.9735. Note, that (3.134) is non-minimum
phase, however both the two stage PE-UIO and the INPEST methods can cope with the
zero at −4.3023. The zero at z0 = 0.9735 needs to be eliminated from the parity equation
in the two stage PE-UIO. The output of the system is subjected to white, zero-mean,
Gaussian noise of the variance 0.003, which means that the standard deviation of the
output measurement noise is equal to approximately 6% of the standard deviation of
the output. The INPEST method is optimised for ν = 0.004, which results in q̊e = 0.2635
and τ̊ = 7. The results of the input reconstruction are presented in Fig. 3.11, whereas
the efficacy in terms of R2T and τ is compared in Table 3.4. It can be noted that both

Table 3.4: Comparison of INPEST and two stage PE-UIO. System has zero at 0.9735

Method R2T [%] τ

INPEST 0.3423 7
2-stage PE-UIO (s = 13) 0.2946 6
2-stage PE-UIO (s = 14) 0.2667 7
2-stage PE-UIO (s = 15) 0.2721 7

algorithms yield comparable results for τ = 7.

3.9 Concluding remarks

An approach to the unknown input reconstruction problem has been proposed. The
scheme is applicable to MIMO systems with a single unmeasurable input, whereas the
number of outputs and known inputs may be arbitrary. The generalised scheme is
suitable for OE and ARMAX linear systems. It is also applicable in the EIV case.
An extension to the standard PE-UIO, namely, the two stage PE-UIO, has been also
proposed. The latter copes with systems which contain a derivative term or whose
zeros lie close to unity (these are cases when the standard PE-UIO is not applicable).
In the numerical study the developed algorithms have been compared with two other

75
3. Parity equations-based unknown input reconstruction for linear stochastic systems

Error
0.4
0.2
0
−0.2
−0.4
0 50 100 150 200 250 300 350 400
Unknown input estimate
4.5 true
4 PE−UIO
INPEST
3.5
3
2.5
2
1.5
1
0.5
0
0 50 100 150 200 250 300 350 400
Time [samples]

Figure 3.11: Comparison of unknown input reconstruction efficacy of the two stage
PE-UIO and the INPEST; in both cases τ = 7. The system has zeros at
−4.3023 and 0.9735.

methods, namely, the Kalman filter-based MVU and the INPEST method which is
based on a closed loop control concept.
The main advantage of the PE-UIO is its simplicity; the filter parameters are cal-
culated once at the beginning of the reconstruction process. The method is fast as it
utilises two moving average filters. The only tuning parameter of the PE-UIO is the
order of the parity space s. By altering it, the bandwidth of the input reconstruction
filter is shaped. This property allows the designer to tune the algorithm for different
levels of noise. It should be noted that by reduction of the filter bandwidth (hence
improvement of the noise filtering properties of the scheme) an estimation lag is intro-
duced. Similar property has been observed in the INPEST method, whereas the MVU
does not allow for introduction of an estimation lag in order to reduce the impact of
the noise (in terms of a bandwidth reduction). Furthermore, the PE-UIO is suitable
for non-minimum phase systems.
The two versions of the PE-UIO algorithm have been compared. The two stage PE-
UIO allows to eliminate selected system zeros from the PE, which changes the response
of the input reconstruction filter. This is particularly desirable, when the system zeros
lie close to unity, which results in a large overshoot in the step response of the standard
PE-UIO. The design of the two stage PE-UIO allows the elimination of this overshoot
and hence a distortion of the unknown input estimate. Furthermore, the two stage
PE-UIO provides better (in terms of bandwidth) noise filtering properties. On the

76
3. Parity equations-based unknown input reconstruction for linear stochastic systems

other hand it has been shown using a numerical example that the phase lead caused
by the system zero when using the standard PE-UIO may be desirable. This might be
the case in an on-line application when the estimation delay is crucial. The phase lead,
when using the standard PE-UIO, result in a reduced estimation delay compared to
the two stage PE-UIO.
The comparison of both PE-UIO methods, the INPEST, and the MVU revealed
comparable efficacy of the PE-UIO and the INPEST. Both algorithms have the pos-
sibility to shape the filter bandwidth (by introducing an inherent delay) by selection
of a single tuning parameter (ν in the case of the INPEST and s in the case of the
PE-UIO). The MVU does not have such a possibility of shaping the bandwidth of the
filter to this extent as in the case of the PE-UIO or the INPEST.
Further work aims towards an extension of the algorithms to systems with multiple
unmeasurable inputs. Although the proposed algorithms are generally applicable for
nonminimum-phase systems, a solution for systems, whose nonmiminum-phase zero is
close to unity, still remains an open question.

77
Chapter 4

Parity equations-based unknown

input reconstruction for
Hammerstein-Wiener systems

Nomenclature

ai . . . . . . . . . . . . . . autoregressive parameter in polynomial model

A ............... state transition matrix in state-space model
bi . . . . . . . . . . . . . . exogenous parameter in polynomial model
B .............. input matrix of known input in state-space model
ci . . . . . . . . . . . . . . moving average parameter in ARMAX model
C .............. output matrix in state-space model
D .............. feedforward matrix of known input in state-space model
e(t) . . . . . . . . . . . . noise term
f (⋅) . . . . . . . . . . . . function to be minimised by Lagrange multiplier method
g(⋅) . . . . . . . . . . . . . constraint function in Lagrange multiplier method
G .............. input matrix of unknown input in state-space model
Gu (z) . . . . . . . . . . z-domain transfer function between u0 (t) and y(t)
Gv (z) . . . . . . . . . . z-domain transfer function between v(t) and y(t)
H .............. feedforward matrix of unknown input in state-space model
m .............. number of system outputs
n ............... order of the system
na . . . . . . . . . . . . . . order of autoregressive polynomial
nb . . . . . . . . . . . . . . order of exogenous polynomial
nc . . . . . . . . . . . . . . order of moving average polynomial
p ............... number of known inputs to a system
pi . . . . . . . . . . . . . . element of P
P .............. auxiliary vector
Q .............. block Toeplitz matrix
s ............... parity space order
S ............... auxiliary matrix
T ............... block Toeplitz matrix
u0 (t) . . . . . . . . . . . noise-free known input

78
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

u(t) . . . . . . . . . . . . measured input

ũ(t) . . . . . . . . . . . . input measurement noise
ū0 (t) . . . . . . . . . . . input to linear dynamic block
ũ∗ (t) . . . . . . . . . . . auxiliary variable
˜(t) . . . . . . . . . . . .
ū estimation error of ū0 (t)
U0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of u0 (t)
U0 (z) . . . . . . . . . . u0 (t) in z-domain
˜ (t) . . . . . . . . . . . .
Ū ˜(t)
stacked vector of last s + 1 values of ū
U (t) . . . . . . . . . . . . stacked vector of last s + 1 values of u(t)
U (z) . . . . . . . . . . . u(t) in z-domain
Ũ (t) . . . . . . . . . . . . stacked vector of last s + 1 values of ũ(t)
Ũ (z) . . . . . . . . . . . ũ(t) in z-domain
Ũ ∗ (t) . . . . . . . . . . . stacked vector of last s + 1 values of ũ∗ (t)
Ũ ∗ (z) . . . . . . . . . . ũ∗ (t) in z-domain
v̂(t) . . . . . . . . . . . . unknown input estimate
v(t) . . . . . . . . . . . . unknown (unmeasurable) input
V (t) . . . . . . . . . . . . stacked vector of last s + 1 values of unknown input
V (z) . . . . . . . . . . . v(t) in z-domain
V̂ (z) . . . . . . . . . . . v̂(t) in z-domain
w qi . . . . . . . . . . . . . auxiliary polynomial parameter
wξi . . . . . . . . . . . . . auxiliary polynomial parameter
W .............. vector, which belongs to Γ–
W (z) . . . . . . . . . . . polynomial of z-variable defined by appropriate elements of W
WQ (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of WQ
WT (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of WT
WT ′ (z) . . . . . . . . . polynomial of z-variable defined by appropriate elements of WT′
WΞ (z) . . . . . . . . . . polynomial of z-variable defined by appropriate elements of WΞ
x(t) . . . . . . . . . . . . state vector instate space model
y0 (t) . . . . . . . . . . . noise-free system output
ȳ0 (t) . . . . . . . . . . . output of linear dynamic block
ȳ˜(t) . . . . . . . . . . . . estimation error of ȳ0 (t)
y(t) . . . . . . . . . . . . measured output
ỹ(t) . . . . . . . . . . . . output measurement noise
Y0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of noise-free output
Ȳ0 (t) . . . . . . . . . . . stacked vector of last s + 1 values of ȳ0 (t)
Y (t) . . . . . . . . . . . . stacked vector of last s + 1 values of measured output
Ȳ˜ (t) . . . . . . . . . . . . stacked vector of last s + 1 values of ȳ˜(t)
Y (z) . . . . . . . . . . . y(t) in z-domain
zi . . . . . . . . . . . . . . system zero
αi . . . . . . . . . . . . . . auxiliary parameter
αi . . . . . . . . . . . . . . auxiliary parameter
δ ............... system delay
ǫ(t) . . . . . . . . . . . . auxiliary noise term
ǫ∗ (t) . . . . . . . . . . . auxiliary noise term
φu (t), φy (t) . . . . . auxiliary variance terms
ϕ(⋅) . . . . . . . . . . . . Hammerstein nonlinearity
γ ............... row vector of Γ–
Γ ............... extended observability matrix
Γ– . . . . . . . . . . . . . . left nullspace of Γ
η(⋅) . . . . . . . . . . . . Wiener nonlinearity

79
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

η −1 (⋅) . . . . . . . . . . inverse of η(⋯)

λ ............... Lagrange multiplier
Π ............... input matrix of noise term in state-space model
Ω ............... feedforward matrix of noise term in state-space model
Σ, Σe , Σũ , Σũe . . covariance matrices
τ ............... unknown input estimation lag
Ξ ............... block Toeplitz matrix
ψ ............... auxiliary vector

Preliminary reading: Sections 2.2, 2.3, 2.4, Subsection 2.5.2, Sections 3.2 and 3.3.

4.1 Introduction
Block oriented models are convenient for modelling nonlinear systems. Their relatively
simple structure of a linear dynamic block interconnected with nonlinear memoryless
function(s) provides a powerful tool for an approximation of a large class of nonlin-
ear systems, see (Pearson & Pottmann 2000, Pearson 2003). Block oriented models
have been used for modelling such phenomena as, for instance: infant EEG (electroen-
cephalogram) seizures (Celka & Colditz 2002), a radio frequency amplifier (Crama &
Rolain 2002), a glucose-insulin process in diabetes type I patient (Bhattacharjee, Sen-
gupta & Sutradhar 2010), ionospheric dynamics (Palanthandalam-Madapusi, Ridley &
Bernstein 2005) or human operator dynamics (Tervo & Manninen 2010). Furthermore,
such models are also used for control purposes, see, for example, (Anbumani, Patnaik
& Sarma 1981, Fruzzetti, Palazoglu & McDonald 1997, De-Feng, Li & Guo-Shi 2010),
and fault detection (Korbicz, Koscielny, Kowalczuk & Cholewa 2003, Lajic, Blanke &
Nielsen 2009).
A two-input single-output Hammerstein-Wiener model is considered, i.e. the linear
dynamic block is preceded and followed by nonlinear static functions. (In the case of a
Hammerstein model a linear block is preceded by a static nonlinear function, whereas
in the case of a Wiener model the order of these elements is reversed.) A problem
of the reconstruction of the unknown/unmeasurable input to the system is taken into
consideration. Up to date, only a limited number of publications are available on this
subject. Szabo, Gaspar & Bokor (2005) proposed an inversion of Wiener systems using
a geometric method based on the assumption that the static nonlinearity transforming
the output is invertible, whilst Ibnkahla (2002) used neural networks for Hammerstein
system inversion.
The algorithm presented here extends the approach developed in Chapter 3 to a
Hammerstein-Wiener case. An EIV framework, see (Söderström 2007), is considered,
i.e. all the measured signals are affected by white, Gaussian, zero-mean and mutually
uncorrelated measurement noise sequences. The theory described in Sections 4.2–4.3
has been presented in (Sumislawska, Larkowski & Burnham 2012).
This chapter is organised as follows: in Section 4.2, for completeness, the idea of

80
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

block oriented models is presented and the Hammerstein-Wiener model, for which the
unknown input reconstruction algorithm is designed, is defined. Then, the PE-UIO
method for Hammerstein-Wiener systems (PE-UIO-HW) is described in Section 4.3.
Furthermore, in Section 4.4 the PE-UIO-HW is extended to an adaptive version, which
accounts for changes in noise levels. Finally, conclusions are provided in Section 4.5.

4.2 Problem statement

It is assumed that a two-input single-output nonlinear system can be described by a
Hammerstein-Wiener model. An EIV framework is considered (Söderström 2007), see
Fig. 4.1. Thus, the Hammerstein-Wiener model is given by the following state-space
form:

ū0 (t) = ϕ (u0 (t)) (4.1a)

x(t + 1) = Ax(t) + B ū0 (t) + Gv(t) (4.1b)
ȳ0 (t) = Cx(t) + Dū0 (t) + Hv(t) (4.1c)
y0 (t) = η (ȳ0 (t)) (4.1d)
u(t) = u0 (t) + ũ(t) (4.1e)
y(t) = y0 (t) + ỹ(t) (4.1f)

where ϕ(⋅) is a static nonlinearity transforming the first system input u0 (t) into an in-
accessible signal ū0 (t) which serves as the first input to the linear block. It is assumed
that the second input v(t) is fed directly (without a nonlinear transformation) to the
linear block, which is described by a state-space model, where A ∈ Rn×n , B ∈ Rn×1 ,
C ∈ R1×n , D ∈ R1×1 , G ∈ Rn×1 and H ∈ R1×1 . The term ȳ0 (t) refers to the output of
the linear part of the system, which is then transformed by the memoryless function
η(⋅) into the overall system output y0 (t). Since the EIV case is considered, all mea-
sured variables, which are u(t) and y(t), are affected by white, Gaussian, zero-mean,
and mutually uncorrelated measurement noise sequences denoted by ũ(t) and ỹ(t), re-
spectively. Noise sequences are postulated to be uncorrelated with the noise-free but
unmeasured system input and output, denoted as u0 (t) and y0 (t), respectively. It is
assumed here that η(⋅) is strictly monotonic, hence its inverse exists. Note that (4.1)
represents a Hammerstein or a Wiener model if, respectively, η(⋅) or ϕ(⋅) is an identity
function.
Similarly as in Chapter 3, the objective of the proposed scheme is to estimate the
unknown input v(t), simultaneously minimising the effect of the measurement noise on
the unknown input estimate. It is assumed that the model of the system is known and
that v(t) is varying relatively slowly, cf. Subsection 3.5.1.

81
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

v(t)

u0 (t) ū0 (t) ȳ0 (t) y0 (t) y(t)

ϕ(u0 ) η(ȳ0 )
linear
system
ỹ(t)
ũ(t) u(t)

Figure 4.1: A Hammerstein-Wiener system in the EIV framework

4.3 PE-UIO for Hammerstein-Wiener systems

In this section the algorithm for estimation of the unknown input is derived. Firstly, for
completeness, the PE for Hammerstein-Wiener systems are derived in Subsection 4.3.1.
This is followed by a development of a PE-UIO-HW algorithm for the considered class
of block oriented systems in Subsections 4.3.2 and 4.3.3.

4.3.1 Parity relations for Hammerstein-Wiener system

Consider the system described by (4.1). Analogously, as in Equation (3.2), one can
build stacked vectors of y(t), y0 (t), ȳ0 (t), ỹ(t), ū0 (t), u(t), u0 (t) and ũ(t) which are
denoted, respectively, as Y (t), Y0 (t), Ȳ0 (t), Ỹ (t), Ū0 (t), U (t), U0 (t) and Ũ (t). By
making use of this notation the system defined by (4.1) can be expressed in the form
of:

Ū0 (t) = ϕ(U0 (t)) (4.2a)

Ȳ0 (t) = Γx(t − s) + QŪ (t)0 + T V (t) (4.2b)
Y0 (t) = η(Ȳ0 (t)) (4.2c)

where ϕ(U0 (t)) is a vector whose elements are ϕ(u0 (t−s)), ϕ(u0 (t−s+1)), ⋯, ϕ(u0 (t)).
Analogously, the function η(Ȳ0 (t)) is defined.
The linear part of the system, defined by (4.2b), can be represented by the following
parity relation, cf. Subsection 3.3.1.

W Ȳ0 (t) = W T V (t) + W QŪ0 (t) (4.3)

which, since η(⋅) is assumed to be invertible, can be reformulated as:

W η −1 (Y0 (t)) = W T V (t) + W Qϕ(U0 (t)) (4.4)

where η −1 (⋅) denotes an inverse of η(⋅). Due to the fact that y0 (t) and u0 (t) are
inaccessible, the parity relation (4.4) can be approximated by the measured values of

82
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

the input and output:

W η −1 (Y (t)) = W T V (t) + W Qϕ(U (t)) + ξ(t) (4.5)

where ξ(t) accounts for an overall error resulting from the presence of measurement
noise. (Note that ξ(t) depends also on the current values of u(t) and y(t) due to the
nonlinearities in the system.) By rearranging the measured (known) variables to the
right-hand side and the unknowns to the left-hand side, the following parity equation
is obtained, cf. (Li & Shah 2002):

W η −1 (Y (t)) − W Qϕ(U (t)) = W T V (t) + ξ(t) (4.6)

4.3.2 Unknown input estimation

Analogously to the PE-UIO it is proposed to estimate the value of the unknown input
as:
v̂(t − τ ) = W η −1 (Y (t)) − W Qϕ(U (t)) (4.7)

which, in the case of noise-free input and output measurements, is:

v̂(t − τ ) = W T V (t) (4.8)

In the case of noisy input and output measurements the unknown input estimate is
affected by an error, cf. (4.6):

v̂(t − τ ) = W T V (t) + ξ(t) (4.9)

resulting from both the input and output measurement uncertainties, which can be
deduced to be given by:

ξ(t) = W (η −1 (Y (t)) − η −1 (Y0 (t))) − W Q (ϕ(U (t)) − ϕ(U0 (t))) (4.10)

Using the notation:

Ȳ˜ (t) = η −1 (Y (t)) − η −1 (Y0 (t))

(4.11)
˜ (t) = ϕ(U (t)) − ϕ(U (t))
Ū 0

Equation (4.10) can be rewritten as:

ξ(t) = W Ȳ˜ (t) − W QŪ

˜ (t) (4.12)

Since ϕ(⋅) and η(⋅) are memoryless, the sequences:

˜(t) = ϕ(u(t)) − ϕ(u0 (t))

ū (4.13)

83
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

and
ȳ˜(t) = η −1 (y(t)) − η −1 (y0 (t)) (4.14)

are white and mutually uncorrelated (as ũ(t) and ỹ(t) are white and mutually uncor-
˜(t), further
related), which is demonstrated further in this Section. The variance of ū
˜(t)), is time varying and depends on ϕ(u(t)), u(t) and the variance
referred to as var(ū
of ũ(t), denoted as var(ũ). Analogously, the variance of ȳ˜(t), i.e. var(ȳ˜(t)), is depen-
dent on var(ỹ) and the current values of η(y(t)) and y(t). The expression ϕ(u0 (t))
can be approximated using a first order Taylor expansion at u(t):

ϕ(u0 (t)) ≈ ϕ(u(t)) + (u0 (t) − u(t)) = ϕ(u(t)) −

∂ϕ(u(t)) ∂ϕ(u(t))
ũ(t) (4.15)
∂u(t) ∂u(t)

˜(t) and ũ(t) can

Thus, incorporating (4.15) into (4.13), the dependency between the ū
be approximated via:

˜(t) = ϕ(u(t)) − ϕ(u0 (t)) ≈

∂ϕ(u(t))
ū ũ(t) (4.16)
∂u(t)

˜(t) and ũ(t) is approximately proportional to the

This means that the ratio between ū
tangential of ϕ(u(t)). Analogously, the ratio between ȳ˜(t) and ỹ(t) is approximately
∂η −1 (y(t))
proportional to ∂y(t) , i.e.

∂η −1 (y(t))
ȳ˜(t) ≈ ỹ(t) (4.17)
∂y(t)

Note that:

˜(t − i)ū
˜(t − j)} ≈ E{
∂ϕ(u(t − i)) ∂ϕ(u(t − j))
E{ū ũ(t − i) ũ(t − j)}
∂u(t) ∂u(t)

}E{ũ(t − i)ũ(t − j)}

∂ϕ(u(t − i)) ∂ϕ(u(t − j))
= E{ (4.18)
∂u(t) ∂u(t)

} × 0 = 0, for i ≠ j
∂ϕ(u(t − i)) ∂ϕ(u(t − j))
= E{
∂u(t) ∂u(t)

Hence, the sequence ū˜(t) is white. Analogously, it can be demonstrated that ȳ˜(t) is
˜(t) and ȳ˜(t) are mutually uncorrelated.
white as well as ū
˜(t) and ȳ˜(t) can be approximated, respectively, as:
The variances of ū

2
˜(t)) ≈ ( ) var(ũ)
∂ϕ(u(t))
var(ū
∂u(t)
(4.19)
∂η −1 (y(t))
2
var(ȳ˜(t)) ≈ ( ) var(ỹ)
∂y(t)

˜(t)) and var(ȳ˜(t)) are, in general, time varying as they

It should be noted that var(ū
depend on the current values of the functions ϕ(u(t)) and η −1 (y(t)). Furthermore, the

84
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

˜(t)) and var(ȳ˜(t)) is not constant, i.e. the impact of either input
ratio between var(ū
or output measurement noise on the unknown input estimation error can be prevailing,
depending on the system operating point. Therefore, the unknown input reconstruction
filter should adapt to these changes.
The aim of the PE-UIO for Hammerstein-Wiener systems is to select such a vector
W that the variance of the error term ξ(t) is minimised, i.e.

var(ξ(t)) = E{(W Ȳ˜ (t) − W QŪ

˜ (t))(W Ȳ˜ (t) − W QŪ
˜ (t))T }
(4.20)
= W Σȳ˜W T + W QΣū˜ QT W T − W ΣTū˜ȳ˜QT W T − W QΣū˜ȳ˜W T

˜ (t)Ū
where Σū˜ = E{Ū ˜ T (t)}, Σ = E{Ȳ˜ (t)Ȳ˜ T (t)}, Σ = E{Ū
˜ (t)Ȳ˜ T (t)}. The term Σ
ȳ˜ ˜ȳ˜
ū ˜
ū
is calculated via, cf. (4.19):

⎡ var(ū
˜(t − s)) ⎤
⎢ ⎥
⎢ ⎥
⋯ 0 0
⎢ ⎥
⎢ ⎥
Σū˜ = ⎢ ⎥
⋮ ⋱ ⋮ ⋮
⎢ ⎥
(4.21)
⎢ ˜(t − 1)) ⎥
⎢ ⎥
0 ⋯ var(ū 0
⎢ var(ū(t)) ⎥
⎣ 0 ⋯ 0 ˜ ⎦

Analogously, the expression Σȳ˜ is obtained by replacing the terms var(ū˜(⋅)) in (4.21)
by var(ȳ˜(⋅)). Due to the fact that ũ(t) and ỹ(t) are mutually uncorrelated, Σū˜ỹ = 0.
For convenience, an expression Σ is introduced, which is equal to:

Σ = Σȳ˜ + QΣū˜ QT (4.22)

Subsequently, the vector W should be selected to minimise the cost function f (W ):

f (W ) = W ΣW T (4.23)

subject to the following constraints:

1. The sum of elements of W T is equal to 1

2. W Γ = 0

The solution to the constrained optimisation problem has been solved using the La-
grange multiplier method in Chapter 3.
˜(t)) and var(ȳ˜(t)) is
Note that due to the fact that the ratio of the variances var(ū
changing over the time, cf. (4.19), as opposed to the linear case in Chapter 3, the vector
W needs to be updated at each time step, i.e. the elements of W are time varying. This
may eventually result in an unnecessary jitter of the estimation lag τ . This happens if
∑ αi i
the mantissa of ∑ αi
, cf. (3.11) and (3.12), is close to 0.5 and in some time instances it
exceeds 0.5, whilst in the other is lower than 0.5. Thus, it is suggested to calculate τ
only once at the beginning of the input reconstruction process. Finally, the algorithm
for calculating the optimal vector W is summarised as follows:

85
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

Algorithm 4.1 (PE-UIO-HW).

1. Select the order of the parity space s ≥ n and build matrices Γ, Q and T .

2. Obtain Γ– .

for t = 1 ∶ N

˜(t) and ȳ˜(t) using (4.19)

3. Calculate variances of ū

4. Compute Σ using:

Σ = Σȳ˜ + QΣū˜ QT (4.24a)

5. Calculate the column vector S via:

S = Γ– Σ(Γ– )T + (Γ– Σ (Γ– )T )

T
(4.24b)

6. Compute the matrix ψ by making use of:

ψ = sumrow (Γ– T ) (4.24c)

7. Obtain the Lagrange multiplier λ as:

λ = ((S −1 ψ) ψ)
T −1
(4.24d)

8. Calculate the parameter vector P by:

P = λS −1 ψ (4.24e)

9. Compute the vector W as:

W = P T Γ– (4.24f)

if t = 1

Calculate the estimation lag as:

∑ αi i
τ = round ( ) (4.24g)
∑ αi

where:
W T = [ αs αs−1 ⋯ α0 ]
T

86
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

end

10. Obtain the estimate of v(t − τ ) via:

v̂(t − τ ) = W (t)η −1 (Y (t)) − W (t)Q(t)ϕ(U (t)) (4.24h)

end

4.3.3 Confidence bounds

It can be seen from (4.20), that the variance of the error term var(ξ(t)) can be repre-
sented as a sum of two terms, each of which depends solely on either the output or the
input measurement noise, such as:

var(ξ(t)) = φu (t) + φy (t) (4.25)

where φu (t) and φy (t) are defined as:

φu (t) = W QΣū˜ QT W T (4.26a)

φy (t) = W Σȳ˜W T
(4.26b)

Therefore, it can be noted that the PE-UIO-HW algorithm minimises the sum of φu (t)
and φy (t).
The accuracy of the unknown input estimation alters over the time, as var(ξ(t)) is
changing. Based on the assumption of a Gaussian distribution of ũ(t) and ỹ(t) it can
be assumed that the distribution of ξ(t) can be approximated with a Gaussian curve
with the variance of var(ξ(t)). Consequently, confidence bounds of the unknown input
estimate can be approximated using Gaussian distribution tables as multiplicities of
the standard deviation of ξ(t).

4.3.4 Numerical examples

Example 4.1. Design of the PE-UIO-HW
Consider an examplary system, whose matrices of the linear block are given by:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ 0 −0.56 ⎥ ⎢ −0.1200 ⎥ ⎢ ⎥
A=⎢⎢ ⎥ ⎢ ⎥ G = ⎢ 0.0055 ⎥
⎥ B=⎢ ⎥ ⎢ ⎥
⎢ 1 1.5 ⎥ ⎢ 0.4125 ⎥ ⎢ 0.0963 ⎥
(4.27)
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
C =[ 0 1 ] D = 0.125 H = 0.025

87
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

The memoryless input and output nonlinearities are arbitrarily selected as:

ū0 (t) = exp (0.165 ⋅ 10−5 u30 (t) + u0 (t)) − 1

(4.28)
y0 (t) = exp (11 + 0.165 ⋅ 10−5 (ȳ0 (t))3 ) − exp(11)

Fig. 4.2 depicts functions ϕ(⋅) and η(⋅) where it is observed that they are both mono-

Hammerstein nonlinearity ϕ(⋅)

2

1.5
ū0 (t)

0.5

u0 (t)
0 0.2 0.4 0.6 0.8 1 1.2

Wiener nonlinearity η(⋅)

80

60
y0 (t)

40

20

ȳ0 (t)
0 2 4 6 8 10

Figure 4.2: Hammerstein and Wiener nonlinearities

tonic and strictly increasing. This means that the impact of the input measurement
noise on the unknown input estimate is expected to be relatively low for low values of
u(t) (as the gradient of ϕ(u(t)) is small for low values of u(t)). On the other hand, this
impact will be relatively high for large values of u(t) (as the gradient of ϕ(u(t)) is large
for high values of u(t)). Due to the fact that the scheme utilises an inversion of η(⋅), an
opposite situation is expected according to the output measurement noise. Low values
of the output are expected to yield a significant impact of the output measurement
error on the accuracy of the unknown input estimate.
The known input and output signals as well as ū0 (t) and ȳ0 (t) are presented in
Fig. 4.3. For the first 1000 samples of the simulation y0 (t) is relatively high and, as
the slope of η(⋅) becomes steeper for higher values of ȳ0 (t), it is anticipated that the
inversion of the noisy measurement y(t) for the first 1000 samples will significantly
reduce the impact of the output measurement noise. After 1000 samples both u0 (t)
and y0 (t) decrease, which results in a higher vulnerability of the input reconstruction
process to the output measurement noise, cf. the slope of η(⋅) for the relatively low
values of the output. The input and output measurements are subjected to white,
Gaussian, zero-mean, and mutually uncorrelated noise sequences, whose variances are,

88
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

2
u0 (t)
u0 (t), ū0 (t) ū0 (t)
1

0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]
y0 (t)
y0 (t), ȳ0 (t)

10
2 ȳ0 (t)

0
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.3: Input and output of the considered system (grey solid curve) compared
with input and output of linear block (dashed-dotted curve)

respectively, var(ũ(t)) = 0.002 and var(ỹ(t)) = 0.5. As a result of the inversion of η(⋅),
needed for the calculation of ȳ0 (t), the ratio of standard deviations of ȳ˜(t) to ȳ0 (t) is
2.2 %. However, as expected, the impact of the measurement noise on the accuracy of
the estimate of ȳ0 (t) changes over time. For the period between 100 and 900 samples
the standard deviation of ȳ˜(t) is equal to 1.4 % of the standard deviation of ȳ0 (t).
Whereas for the period between 1100 and 1900 samples this ratio is 12.3 %. This
can be interpreted that the impact of the measurement noise decreases over 8 times
after 1000 samples. The order of the parity space has been selected as 12, which gives
τ = 6 samples. The unknown input estimate with 95% confidence bounds is presented
˜(t)) and φu (t) are compared, whilst
in Fig. 4.4. In the upper subfigure of Fig. 4.5 var(ū
the middle subfigure of Fig. 4.5 compares var(ȳ˜(t)) and φy (t). The lower subfigure
of Fig. 4.5 presents the optimisation effect by comparing the sum of var(ū˜(t)) and
var(ȳ˜(t)) with the sum of φu (t) and φy (t). During the first 1000 samples the input
measurement noise has a larger influence on the unknown input estimation error in
comparison to the output measurement noise. One can note that for the first 300
samples the effect of the output measurement noise is actually amplified (as a result
of the minimisation of the joint impact of the input and output measurement noise).
˜(t), it has a negligible effect on the input estimation
However, due to a relatively large ū
error. After 1000 samples of the simulation the situation changes. The effect of the
input measurement noise becomes less significant, whereas the term ȳ˜(t) increases as
it depends strongly on the value of the output.

Example 4.2. Distribution of ξ(t)

In order to calculate confidence bounds of the unknown input estimate a Gaussian
distribution of ξ(t) is assumed. In this example a Monte-Carlo simulation with 10000
runs is carried out and the theoretical distribution of ξ(t) is compared with an experi-

89
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

4.5
confidence bounds
true input
input estimate
4
Unknown input estimate

3.5

2.5

0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.4: Unknown input estimation for Hammerstein-Wiener system in the EIV
framework

Influence of input measurement error

−2
10

˜(t))
var(ū
−3
φu (t)
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Influence of output measurement error
var(ȳ˜(t))
−2 φy (t)
10

−4
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Overall influence of measurement error
˜(t)) + var(ȳ˜(t))
−1
10
var(ū
φu (t) + φy (t)
−2
10

0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.5: Adaptive minimisation of the effect of measurement noise on the input
estimate

mentally obtained probability density function of the variable:

W Ȳ˜ (t)Ȳ˜ T (t)W T + W QŪ

˜ (t)Ū
˜ T (t)QT W T (4.29)

90
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

at each time sample. The system from the previous example is used for the simulation.
In Fig. 4.6 those two distributions have been compared as functions of time. Further-

Figure 4.6: Experimental and calculated distributions of ξ(t) as functions of time.

The coloured surface presents the theoretical distribution of ξ(t) (Gaus-
sian curve with the variance defined by (4.20)) for a single simulation run.
Black plots are experimentally obtained probability density functions of
ξ(t) from 10000 runs of the Monte-Carlo simulation.

more, in Fig. 4.7 both theoretical and experimental distributions of ξ(t) are presented
for four different time instances. It can be noted that the experimentally obtained dis-
tribution of ξ(t) matches the theoretical Gaussian distribution with the variance given
by equation (4.20). In Fig. 4.8 values of φu (t) and φy (t) as functions of time for a single
simulation run have been compared with functions of time of mean values of, respec-
˜ (t)Ū
tively, W QŪ ˜ T (t)QT W T and W Ȳ˜ (t)Ȳ˜ T (t)W T from the Monte-Carlo simulation.
It can be noted that the values of φu (t) and φy (t) calculated using (4.26) match the
experimental data.

Example 4.3. Use of the linear PE-UIO instead of the PE-UIO-HW

As filter parameters are recalculated at each time sample, the PE-UIO-HW algorithm
becomes computationally demanding. Therefore, it is worth considering to approximate
the Hammerstein-Wiener system with a linear model and then use the linear PE-UIO
described in Chapter 3 instead. Use of the linear PE-UIO in a Hammerstein-Wiener
case is, however, feasible only if the nonlinearities are mild enough, so the error resulting
from a linear approximation of the nonlinear system is relatively small. In this example
use of the linear PE-UIO instead of the PE-UIO-HW is considered and a degradation
of performance resulting from the use of the linear algorithm is examined.

91
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

t = 400, φu (t) = 8.3e-3, φy (t) = 3.5e-4 t = 1060, φu (t) = 1.9e-3, φy (t) = 2.3e-3
5 8
histogram of ξ(t)
4 Gaussian6 curve

3
4
2
2
1

0 0
−0.4 −0.2 0 0.2 0.4 −0.2 −0.1 0 0.1 0.2

t = 1280, φu (t) = 1.3e-3, φy (t) = 8.0e-4 t = 1800, φu (t) = 1.7e-3, φy (t) = 2.6e-3
10 8

8
6
6
4
4
2
2

0 0
−0.2 −0.1 0 0.1 0.2 −0.2 −0.1 0 0.1 0.2 0.3

Figure 4.7: Comparison of theoretical distribution of ξ(t) with experimental data.

Black solid lines present theoretical distributions of ξ(t) for different time
instances (Gaussian curves with variance defined by (4.20)) for a single
simulation run. Grey stems are experimentally obtained probability den-
sity functions of ξ(t) from 10000 runs of Monte-Carlo simulation.

The linear block of the considered system is given by (4.28), whilst the Hammerstein
and Wiener nonlinearities are:

ū0 (t) =
10
−5
1 + e−0.4u0 (t) (4.30)
y0 (t) =
bi
+ ci
1 + e i ȳ0 (t)
−a

where ai , bi , and ci are the coefficients of the Wiener nonlinearity η(⋅). The experi-
ment has been performed for three different Wiener nonlinearities (i = 1, 2, 3), whose
coefficients are given in Table 4.1. The Hammerstein nonlinearity as well as the three
considered Wiener nonlinearities, denoted as η1 (⋅), η2 (⋅), and η3 (⋅), are presented in
Fig. 4.9.
Both known and unknown inputs to the system, u0 (t) and v(t), are the same as
in Example 4.1. The upper subfigure of Fig. 4.10 presents u0 (t) and ū0 (t). Due to
the fact that the Hammerstein nonlinearity at the operating point is negligible ū0 (t)

92
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

φu (t)
0.015
obtained from MC
calculated (single run)
0.01

0.005

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000

0
φy (t)
10

−2
10

−4
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.8: Comparison of calculated and experimental values of φu (t) and φy (t).
Black dashed curves present theoretical values of φu (t) and φy (t) as func-
tions of time (calculated using (4.26) for a single simulation run). Grey
curves are experimentally obtained values of φu (t) and φy (t) from 10000
runs of Monte-Carlo simulation.

Table 4.1: Coefficients of Wiener nonlinearities η1 (⋅), η2 (⋅), and η3 (⋅)

i ai bi ci
1 0.25 24 −12
2 0.2 26 −13
3 0.1 43 −21.5

is very close to u0 (t) (the considered system is virtually a Wiener system). The lower
subfigure of Fig. 4.10 shows ȳ0 (t) and the corresponding y0 (t) for three different Wiener
nonlinearities. It is anticipated that the accuracy of the unknown input estimation using
the linear PE-UIO will depend on the severity of the Wiener nonlinearity, i.e. the best
accuracy is expected for η3 (⋅), whilst it is anticipated that η1 (⋅) will result is the most
distorted unknown input estimate. The measured input and the output of the system
are subjected to white, zero-mean, Gaussian, mutually uncorrelated sequences with the
variances, respectively, var(ũ(t)) = 0.002 and var(ỹ(t)) = 0.003.
For each case of a nonlinear system (i.e. a system with different Wiener nonlinearity)
a linear model is obtained using the least squares technique in order to estimate the
unknown input using the PE-UIO with s = 12 samples. A Monte-Carlo simulation with
100 runs is carried out, whose results in terms of R2T are compared with results of the
PE-UIO-HW and presented in Table 4.2 . Sample plots of the unknown input estimate

93
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

Hammerstein nonlinearity ϕ(⋅)

2

1
ū0 (t)
0

−1
u0 (t)
−0.5 0 0.5 1 1.5

Wiener nonlinearities
10
y0 (t)

5 η1 (⋅)
η2 (⋅)
0 η3 (⋅)
−2 0 2 4 6 8 10 12
ȳ0 (t)

Figure 4.9: Hammerstein and Wiener nonlinearities

1
u0 (t), ū0 (t)

u0 (t)
0.5 ū0 (t)

−0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples] ȳ0 (t)
y0 (t) (η1 (⋅))
y0 (t) (η2 (⋅))
y0 (t), ȳ0 (t)

10
y0 (t) (η3 (⋅))
8
6
4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.10: The upper subfigure shows the input of the considered system (grey solid
curve) compared with the input of linear block (dashed-dotted curve).
The lower subfigure presents the output of the linear dynamic block
(grey solid curve) compared with the output of the system for different
Wiener nonlinearities (black dashed, dashed-dotted and dotted curves).

for the considered models are plotted in Fig. 4.11. As expected the distortion in the
unknown input estimate using the linear PE-UIO is least when the Wiener nonlinearity
is given by η3 (⋅), whilst for η1 (⋅) the reconstructed signal is least accurate.

94
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

η1 (⋅)
5
true input
4 PE−UIO−HW
3 PE−UIO

2
Unknown input estimate

0 200 400 600 800 1000 1200 1400 1600 1800 2000
η2 (⋅)
5
4
3
2
0 200 400 600 800 1000 1200 1400 1600 1800 2000
η3 (⋅)
4

2
200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.11: Unknown input estimation for different Wiener nonlinearities

Table 4.2: Comparison of efficacy (in terms of the mean value of R2T [%] from a Monte-
Carlo simulation with 100 runs) of the linear PE-UIO and the PE-UIO-HW
for three different Wiener nonlinearities

η1 (⋅) η2 (⋅) η3 (⋅)

linear nonlinear linear nonlinear linear nonlinear
0.7130 0.0385 0.4349 0.0270 0.0658 0.0216

4.4 Adaptive order PE-UIO for Hammerstein-Wiener sys-

tems
As it has been demonstrated in Subsection 3.7.1 an increase of the parity space order
s reduces the bandwidth of the unknown input reconstructor thus improving the noise
filtering properties of the filter (i.e. reducing the impact of the noise on the unknown
input estimate). However, the reduction of the filter bandwidth results in the input
˜(t)) and var(ȳ˜(t)) are
reconstruction filter being sluggish. Due to the fact that var(ū
time varying, the impact of the noise on the unknown input varies. Therefore, it is
beneficial to vary the bandwidth of the filter (via changing the value of the parity
space order s) as values of var(ū˜(t)) and var(ȳ˜(t)) change. In the algorithm proposed
˜(t))
in this section the order of the parity space varies according to the changes of var(ū
and var(ȳ˜(t)). In order to recognise that the order of the parity space is time varying,
its value at the time instance t is further denoted as s(t).

95
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

4.4.1 Choice of s(t)

˜(t)) and var(ȳ˜(t)). Considering (4.25)
The choice of s(t) should depend on both var(ū
and (4.26) it should be noted that the input and output noise filtering indices defined
as:

φu
˜(t))
ρu = (4.31a)
var(ū
φy
var(ȳ˜(t))
ρy = (4.31b)

are not equal due to the presence of the matrix Q in (4.26a) and hence its influence
on (4.31a). This means that the impact of the change of s(t) will be different for the
input and the output measurement noise. It is proposed to create a two-dimensional
˜(t)) and var(ȳ˜(t)). Fur-
map, which assigns the value of s(t) for each couple of var(ū
˜(t)) and var(ȳ˜(t)) are calculated based on the current
thermore, as the values of var(ū
values of the measured input and output signals (affected by noise), cf. (4.19), the order
˜(t)) and var(ȳ˜(t))
of the parity space s(t) selected based on the current values of var(ū
may jitter unnecessarily. In order to avoid this problem, it is proposed to use local
˜(t)) and var(ȳ˜(t)) defined as:
mean values of var(ū

t+t2
˜(t)) = ∑ (var(ū
˜(i))
1
var(ū (4.32a)
t1 + t2 + 1 i=t−t1
t+t2
var(ȳ˜(t)) = ∑ (var(ȳ˜(i))
1
(4.32b)
t1 + t2 + 1 i=t−t1

where t1 and t2 are arbitrarily defined by the user.

4.4.2 Variable estimation lag

At the time instance t, the following delayed unknown input estimate is calculated:

v̂(t − τ (t)) = W (t)η −1 (Y (t)) − W (t)Q(t)ϕ(U (t)) (4.33)

where τ (t) is time varying, due to the alternating value of s(t). (Note that the notation
τ (t), W (t), and Q(t) has been used instead of τ , W , and Q in order to indicate that
the estimation lag τ , the vector W , and the matrix Q as well as sizes of W and Q are
time varying.) This would eventually lead to difficulties, such as some time instances
of the unknown input would be omitted, and some of them estimated more than once.
Therefore, a logic must be implemented, which copes with the variable estimation lag.
A difficulty may arise in two situations:

(i) τ (t) > τ (t − 1)

(ii) τ (t) < τ (t − 1)

96
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

In the first case a particular time instance of the unknown input estimate is calculated
twice. In such a case from the two values of the unknown input estimate sample the
one should be selected, which is less affected by the noise. The fact that τ (t) increases,
means an increase of the noise influence, i.e. var(ū ˜(t)) or var(ȳ˜(t)) has increased.
Therefore, the impact of the measurement noise on the unknown input estimate has
also increased. Consequently, it can be deduced that the previously calculated value of
the unknown input estimate is less affected by noise.
In the second case, the situation is opposite, i.e. some time instances of v̂(t) will
be omitted. It is proposed to use W (t − 1) and Q(t − 1) to calculate the missing values
of the unknown input estimate.
Incorporating this logic into Algorithm 4.1 the adaptive order PE-UIO-HW (AO-
PE-UIO-HW) is obtained:

Algorithm 4.2 (AO-PE-UIO-HW).

for t = 1 ∶ N

˜(t)) and var(ȳ˜(t))

● Calculate var(ū
˜(t)) and var(ȳ˜(t)) select s(t)
● Based on var(ū
● Obtain W (t), Q(t), and τ (t) as in Algorithm 4.1
if τ (t) = τ (t − 1)
● Calculate v̂(t − τ (t)) as:
v̂(t − τ (t)) = W (t)η −1 (Y (t)) − W (t)Q(t)ϕ(U (t)) (4.34)

elseif τ (t) < τ (t − 1)

for k = τ (t − 1) ∶ τ (t) − 1

v̂(t−k) = W (t−1)η −1 (Y (t−k))−W (t−1)Q(t−1)ϕ(U (t−k)) (4.35)

end
v̂(t − τ (t)) = W (t)η −1 (Y (t)) − W (t)Q(t)ϕ(U (t)) (4.36)

else
● Do nothing
end

97
 ˜(t)) and var(ȳ˜(t)). The row of the table is
Table 4.3: The table assigns value of the parity space order s(t) based on the values of var(ū
˜(t))} ≤ uu , whereas the column is chosen such that yd < log {var(ȳ˜(t))} ≤ yu
selected such that ud < log {var(ū

4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener

log {var(ȳ˜(t))}
yd −∞ -2.87 2.66 -2.52 -2.42 -2.31 -2.19 -2.10 -2.01 -1.94 -1.88 -1.83 -1.76 -1.68 -1.61
HH y
ud HH u -2.87 -2.66 -2.52 -2.42 -2.31 -2.19 -2.10 -2.01 -1.94 -1.88 -1.83 -1.76 -1.68 -1.61 ∞
uu HH
−∞ -3.3 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
-3.3 -3.2 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
-3.2 -3.0 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
-3.0 -2.9 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
-2.9 -2.8 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
-2.8 -2.6 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
˜(t))}

-2.6 -2.5 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
log {var(ū

-2.5 -2.3 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
-2.3 -2.1 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
-2.1 -2.0 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
-2.0 -1.8 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
-1.8 -1.7 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
-1.7 -1.6 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
-1.6 -1.4 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
-1.4 -1.1 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

systems
-1.1 -0.7 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
98
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

4.4.3 Numerical examples

In this subsection the design of the AO-PE-UIO-HW is presented. Two examples are
considered here. The first example resembles an OE case, i.e. the input measurement
is noise-free whilst the output is subjected to white, Gaussian, zero-mean measurement
noise. The second example is in the EIV framework, i.e. both the system input and
the output are affected by white, Gaussian, zero-mean measurement noise.

Example 4.4. Design of the AO-PE-UIO-HW in an OE noise case

Consider a system defined by (4.27) and (4.28). It is assumed that the output of
the system is subjected to white, Gaussian, zero-mean noise sequence of the variance
var(ỹ) = 0.5, whereas var(ũ) = 0 (OE case).
The input and output signals as well as ū0 (t) and ȳ0 (t) are presented in Fig. 4.12.
Similarly as in Example 4.1, y0 (t) is relatively high for the first 1000 samples of the
simulation, hence it is anticipated that the inversion of the noisy measurement y(t) for
the first 1000 samples will reduce the impact of the output measurement noise. This is
due to relatively steep slope of η(⋅) for high values of y0 (t). After 1000 samples y0 (t)
decreases, which is expected to result in a higher vulnerability of the input reconstruc-
tion process to the output measurement noise, as the slope of η(⋅) is less steep for the
relatively low values of the output.

2
u0 (t)
u0 (t), ū0 (t)

ū0 (t)
1

0 200 400 600 800 1000 1200 1400 1600 1800 2000

2 y0 (t)
y0 (t), ȳ0 (t)

10
ȳ0 (t)

0
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.12: Input and output of the considered system (grey solid curve) compared
with input and output of linear block (dashed-dotted curve)

The unknown input in this example is slightly lower than that in Example 4.1, which
yields lower values of y0 (t) compared to Example 4.1. Therefore, it is anticipated that
the effect of measurement noise on the unknown input reconstruction process will be
more significant than in Example 4.1 (especially when y0 (t) is very low between 1600
and 1800 sample).
As a result of the inversion of η(⋅), needed for the calculation of ȳ0 (t), the ratio of

99
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

standard deviations of ȳ˜(t) to ȳ0 (t) is 4.5 %. However, this ratio changes over the time.
For the period between 100 and 800 samples the standard deviation of ȳ˜(t) is equal
to 1.7 % of the standard deviation of ȳ0 (t). Whereas for the period between 1100 and
1600 samples this ratio is 27.0 %. In the extreme case of the period between 1600 and
1800 samples this ratio is equal to 81.1 %. Such a large deviation of the measurement
noise impact requires adaptivity of the unknown input reconstruction scheme. The
term var(ȳ˜(t)) has been calculated with t1 = 2τ (t − 1) + 1 and t2 = 0, cf. (4.32).

4
confidence bounds
true input
Unknown input estimate

3.5 input estimate

2.5

1.5

1
200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.13: Unknown input estimation for a Hammerstein-Wiener OE system using

AO-PE-UIO-HW

Influence of output measurement error

var(ȳ˜(t))
0
10
φy (t)

−5
10
200 400 600 800 1000 1200 1400 1600 1800 2000
Parity space order
30
s(t)
20

10
200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.14: Adaptive minimisation of the effect of measurement output noise on the
input estimate

The unknown input estimate with 95 % confidence bounds is presented in Fig. 4.13.
The parity space order varies according to Table 4.3 and as a function of time is

100
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

presented in the lower subfigure of Fig. 4.14. The upper subfigure of Fig. 4.14 presents
var(ȳ˜(t)) and φy (t). As expected, the parity space order is low in the first half of
the simulation, when the impact of the measurement noise is low. Such a small s(t)
ensures a high bandwidth of the filter, and therefore even high frequency components
of v(t) are reconstructed, cf. Fig. 4.13. In the second half of the simulation, when the
impact of the measurement noise becomes more significant, the order of the parity space
increases. Furthermore, a higher parity space order yields stronger noise attenuation
(in terms of the ratio between var(ȳ˜(t)) and φy (t)), which can be seen in the upper
subfigure of Fig. 4.14.
A Monte-Carlo simulation with 100 runs has been carried out to compare the per-
formance of the AO-PE-UIO-HW and the PE-UIO-HW with a constant parity space
order for two cases of s. The aim of this experiment is to quantify the improvement of
the unknown input reconstruction process when the AO-PE-UIO-HW is used instead
of the PE-UIO-HW. Results in terms of the R2T are compared in Table 4.4. It can be
noted that by varying the parity space order an improvement of the accuracy of the
algorithm has been achieved. However, it needs to be remembered that the adaptive
algorithm needs more computational power.

Table 4.4: Comparison of efficacy of the PE-UIO-HW and the A0-PE-UIO-HW in

terms of R2T [%]

PE-UIO-HW AO-PE-UIO-HW
sample s = 10 s = 23 s = 26 variable s
100:1990 0.0305 0.0320 0.0438 0.0197
100:1000 6.5e-4 0.0023 0.0033 5.7e-4
1000:1990 0.0600 0.0609 0.0837 0.0378

Example 4.5. Design of the AO-PE-UIO-HW in the EIV framework

Similarly as in Example 4.4 the Hammerstein-Wiener system is defined by (4.27)
and (4.28). The input and output signals as well as ū0 (t) and ȳ0 (t) are the same
as in Example 4.4, cf. Fig. 4.12. Both input and output measurements are subjected
to white, Gaussian, zero-mean, mutually uncorrelated noise sequences, whose variances
are, respectively, var(ũ) = 0.001 and var(ỹ) = 0.5, i.e. EIV framework. Similarly as
in Example 4.5, for the first 1000 samples the output signal is relatively high, hence
the output measurement error is expected to have a relatively low impact on the input
reconstruction error. However, as the known input is relatively high for the first half
of the simulation, whilst the slope ϕ(⋅) is relatively small, it is anticipated that the im-
pact of the input measurement noise on the estimation error will be prevailing for the
first 1000 samples of the simulation, cf. Example 4.1. In contrast, after 1000 samples,
when both u0 (t) and y0 (t) decrease, the influence of the output measurement noise is
expected to increase, whilst the impact of the input measurement noise is anticipated

101
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

to reduce. Similarly as in the previous example the parity space order has been ob-
˜(t)) and var(ȳ˜(t)) have been calculated
tained using Table 4.3, whereas terms var(ū
using t1 = 2τ (t − 1) + 1 and t2 = 0.

4
confidence bounds
true input
Unknown input estimate

3.5 input estimate

2.5

1.5

1
200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.15: Unknown input estimation for a Hammerstein-Wiener system in the EIV
framework using AO-PE-UIO-HW

The unknown input estimate is presented in Fig. 4.15. The values var(ū ˜(t)) and
φu (t)) are depicted in the upper subfigure of Fig. 4.16, whilst var(ȳ˜(t)) and φy (t)), are
shown in the middle subfigure of Fig. 4.16. The lower subfigure of Fig. 4.16 presents
the parity space order s(t) as a function of time, which is compared with the s(t) from
the previous example. Note that the only difference between Examples 4.4 and 4.5 is
presence of the input measurement noise. The term var(ū˜(t)) is relatively large for
the first half of the simulation, whereas it becomes negligible after 1000 samples. This
influence of the input measurement noise can be noticed by comparing the values of
s(t) for the two considered examples. The presence of the input measurement noise
causes an increase of s(t) by approximately 5 samples compared to the OE case during
the first half of the simulation. After the first 1000 samples, as the impact of the input
measurement noise on the unknown input estimate becomes negligible, s(t) is similar
for both the OE (Example 4.4) and the EIV (Example 4.5) cases.

4.5 Concluding remarks

The algorithms presented in this chapter are extensions of the PE-UIO developed in
Chapter 3. The basic idea of the unknown input reconstruction scheme for Hammerstein-
Wiener systems is to, firstly, knowing the system nonlinearities calculate the known in-
put and the output of the linear block, then use the PE-UIO to calculate the unknown
input. The algorithm has been developed for the EIV framework, i.e. when both mea-
sured input and output of the system are subjected to white, Gaussian, zero-mean,

102
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

Influence of output measurement error

var(ȳ˜(t))
φy (t)
0
10

−2
10

−4
10
200 400 600 800 1000 1200 1400 1600 1800 2000

−2
Influence of input measurement error
10

−3

˜(t))
10
var(ū
φu (t)
200 400 600 800 1000 1200 1400 1600 1800 2000
Parity space order
Example 4.5
25
Example 4.4
20

15

10
200 400 600 800 1000 1200 1400 1600 1800 2000
Time [samples]

Figure 4.16: Upper and middle subfigures demonstrate adaptive minimisation of the
effect of measurement output noise on the input estimate in Example 4.5.
Lower subfigure compares values of parity space order s(t) as functions
of time in Examples 4.4 (OE) and 4.5 (EIV).

mutually uncorrelated noise sequences. The calculated values of the known input and
the output of the linear block are affected by measurement noise and the impact of
the EIV disturbance sequences depends on the values of the known input and out-
put themselves. This is due to the nonlinearities preceding and following the linear
block. Consequently, the impact of the measurement noise on the unknown input es-
timate changes over time. Therefore, the filter parameters are calculated at each time
instance.
In the first of the proposed algorithms, the PE-UIO-HW, the order of the parity
space, and thus the estimation delay, remains constant, whilst the filter parameters
vary over time.
Due to the fact that the impact of the measurement noise on the unknown input
estimate may vary significantly, a further extension to the scheme is proposed, where the
order of the parity space, s(t), is time varying. The parity space order is selected based
on values of the input and output measurement noise impact coefficients, calculated
as functions of noise variances and measured signals. The variable parity space order
allows the adjustment of the bandwidth of the unknown input reconstruction filter

103
 4. Parity equations-based unknown input reconstruction for Hammerstein-Wiener
systems

to the changing impact of the measurement noise on the unknown input estimate.
The variation of s(t) imposes a variable estimation lag, τ (t). Consequently, a logic is
implemented, which resolves the problem of time varying τ (t), resulting in a smooth
estimate of the unknown input.
The proposed schemes, since inherently adaptive, require at each discrete time
step a non negligible computational effort. The future work, therefore, aims towards
an optimisation of the computational procedure. It is also intended to extend the
algorithm to the multivariable case. Furthermore, block oriented models in the EIV
framework are a new topic in the literature, for which effective identification schemes are
required. Although it has been assumed that the unknown input is fed directly to the
linear block, the algorithm can be easily extended to the case, when the unknown input
is transformed by a nonlinear memoryless function, and afterwards, the transformed
unknown input is fed to the linear dynamic block (based on the assumption that the
static nonlinearity is invertible).

104
Chapter 5

Robust fault detection via

eigenstructure assignment

Nomenclature

A ............... state transition matrix in state-space model

Ac1 , Ac2 . . . . . . . . filter state transition matrices
(i)
Aλ , A∗e , A∗e , Aw auxiliary matrices
˜
Ã, Ã . . . . . . . . . . . . auxiliary matrices
′
A .............. auxiliary matrix
B .............. input matrix of the input in state-space model
˜ ..............
B̃ auxiliary matrix
C .............. output matrix in state-space model
C′ . . . . . . . . . . . . . . auxiliary matrix
d(t) . . . . . . . . . . . . disturbance signal
di (t) . . . . . . . . . . . ith element of d(t)
d∗ (t) . . . . . . . . . . . d(t), whose elements di (t) are delayed, respectively, by δi
D .............. feedforward matrix of known input in state-space model
e ............... matrix of directions of elements of d(t)
ei . . . . . . . . . . . . . . direction of di
ē . . . . . . . . . . . . . . . matrix built from matrices ei
ēi . . . . . . . . . . . . . . matrix whose image is sum of image of ei and images of invariant zero directions
of (A, ei , C)
E .............. input matrix of disturbance signal in state-space model
Ei . . . . . . . . . . . . . . ith column of input matrix of disturbance signal in state-space model
F .............. input matrix of fault signal in state-space model
F (ē) . . . . . . . . . . . projection of F on subspace spanned by columns of ē
F (ē– ) . . . . . . . . . . projection of F on subspace orthogonal to ē
g, gi . . . . . . . . . . . . auxiliary scalar
I ............... identity matrix
J, J1 , J2 . . . . . . . . . gain matrices
K, K1 , K2 , K ′ . . . gain matrices
lj , lj∗ . . . . . . . . . . . . transposes of left eigenvectors of filter state transition matrix
m .............. number of system outputs
n ............... number of states in state-space model

105
 5. Robust fault detection via eigenstructure assignment

p ............... number of system inputs

P (λi ) . . . . . . . . . . auxiliary function of λi
P, R . . . . . . . . . . . . auxiliary matrices
Pl,k . . . . . . . . . . . . auxiliary term
q ............... number of disturbance signals
Q .............. gain matrix
r ............... number of fault signals
r(t) . . . . . . . . . . . . residual
ri (t) . . . . . . . . . . . . ith element of r(t)
rq . . . . . . . . . . . . . . number of rows of Q
T ............... similarity transformation matrix
T1 , T2 . . . . . . . . . . . submatrices of T
u(t) . . . . . . . . . . . . system input
u0 (t) . . . . . . . . . . . noise-free output in output-error case
ũ(t) . . . . . . . . . . . . input measurement noise
U (t) . . . . . . . . . . . . stacked vector of last τ + 1 values of u(t)
U0 (t) . . . . . . . . . . . stacked vector of last τ + 1 values of u0 (t)
Ũ (t) . . . . . . . . . . . . stacked vector of last τ + 1 values of ũ(t)
(i)
v, vj , vj . . . . . . . . auxiliary vectors
Ve . . . . . . . . . . . . . . matrix whose columns are eigenvectors of filter state transition matrix
(i) ′(i)
wj , wj , wj ... right eigenvectors of filter state transition matrix
∗
wj . . . . . . . . . . . . . auxiliary vector
W .............. auxiliary matrix
Wu , Wy . . . . . . . . . parity matrices
x(t) . . . . . . . . . . . . state vector instate space model
x̂(t) . . . . . . . . . . . . state estimate
xi,j . . . . . . . . . . . . . auxiliary scalars
X .............. auxiliary matrix
y(t) . . . . . . . . . . . . system output
y0 (t) . . . . . . . . . . . noise-free output in output-error case
ỹ(t) . . . . . . . . . . . . output measurement noise
Y (t) . . . . . . . . . . . . stacked vector of last τ + 1 values of y(t)
Y0 (t) . . . . . . . . . . . stacked vector of last τ + 1 values of y0 (t)
Ỹ (t) . . . . . . . . . . . . stacked vector of last τ + 1 values of ỹ(t)
zi . . . . . . . . . . . . . . system zero
zi (t) . . . . . . . . . . . . state estimate
αi . . . . . . . . . . . . . . auxiliary parameter
βi , β̄i . . . . . . . . . . . auxiliary parameter vector
δi . . . . . . . . . . . . . . auxiliary term
(i)
λj , λj . . . . . . . . . eigenvalue of filter state transition matrix
Λe . . . . . . . . . . . . . . diagonal matrix whose diagonal elements are eigenvalues of filter state transition
matrix
µ(t) . . . . . . . . . . . . fault signal
µi (t) . . . . . . . . . . . ith element of fault signal
Θ(i) , Θ̄(i) . . . . . . . auxiliary matrices
Ω ............... set of all invariant zeros of (A, e, C) or an auxiliary matrix
Ωi . . . . . . . . . . . . . . set of all invariant zeros of (A, ei , C)
Σũ , Σỹ , Σµ̄ . . . . . . covariance matrices
τ ............... convergence time of finite time-convergent state observer, order of parity space
ξ(t) . . . . . . . . . . . . state estimation error

106
 5. Robust fault detection via eigenstructure assignment

Ψ .............. auxiliary matrix

Preliminary reading: Sections 2.2, 2.6, and 2.7.

5.1 Introduction
Increasing complexity of industrial systems leads to a growing demand for system fault
diagnosis. Furthermore, system uncertainties (disturbances), such as modelling errors,
parameter variations or unmeasurable external stimuli, obstruct the fault detection
process, leading to false alarms. Therefore, a need arises for robust, i.e. disturbance
decoupled, fault detection schemes. In this chapter robust fault detection is consid-
ered. This means that the residual generator is sensitive to faults but insensitive to
disturbances.
Frank & Wünnenberg (1989), Duan & Patton (2001), and Edelmayer (2005) pre-
sented robust fault detection schemes based on unknown input observers. LMI have
been also used for the robust fault detection (Chen & Nagarajaiah 2007, Ding, Zhong,
Bingyong & Zhang 2001). Zhong, Ding, Lam & Wang (2003) proposed an LMI ap-
proach to design a robust fault detection filter for uncertain linear time-invariant sys-
tems. Patton and Chen (Patton & Chen 1991b, Chen & Patton 1999) used the left
and right eigenstructure assignment techniques for the purpose of disturbance decou-
pling. Furthermore, the equivalence between the left eigenstructure assignment-based
robust fault detection filter and the first order PE has been demonstrated by Patton &
Chen (1991a, 1991b, 1991c). Also the problem of the robust fault detection via eigen-
structure assignment has been of the topic of the research of Park & Rizzoni (1994)
and Shen & Hsu (1998). Douglas & Speyer (1995) proposed an algorithm for pre-
venting ill-conditioning when using left eigenstructure assignment. A novel method for
left eigenstructure assignment has been proposed in (Kowalczuk & Suchomski 2005).
Patton & Liu (1994) presented a robust control design method using eigenstructure
assignment, genetic algorithms and a gradient-based optimisation. A reconfigurable
control scheme has been presented in (Ashari, Sedigh & Yazdanpanah 2005a, Ashari,
Sedigh & Yazdanpanah 2005b).
Eigenstructure assignment has been used in various industrial applications. A re-
view of applications has been presented in (Isermann & Balle 1997). Robust fault
detection filters based on eigenstructure assignment have been used in a rolling mill
(Gu & Poon 2003), a jet engine (Patton & Chen 1992), an automotive engine (Shen &
Hsu 1998), an advanced vehicle control systems (Douglas, Speyer, Mingori, Chen, Mal-
ladi & Chung 1996), a single-shaft gas turbine (Fantuzzi, Simani & Beghelli 2001), a
flexible manipulator (Tan & Habib 2006), an inverted pendulum (Tan & Habib 2004), a
vehicle health monitoring system (Ng, Chen & Speyer 2006), and a longitudinal motion
of an unmanned aircraft model (Siahi, Sadrnia & Darabi 2009). Luenberger state ob-
servers using a fixed-structure H∞ optimization have been applied to fault detection of a

107
 5. Robust fault detection via eigenstructure assignment

lane-keeping control of automated vehicles (Ibaraki, Suryanarayanan & Tomizuka 2005)

and fifth order linearised dynamics of an aircraft (Ashari et al. 2005a).
This chapter is organised as follows: in Section 5.2 the robust fault detection filter
design method via right eigenstructure assignment of Chen & Patton (1999) is extended
to systems whose output response to disturbances contains invariant zeros. Then design
of robust PE via right eigenstructure assignment is proposed in Section 5.3. Further-
more, in Section 5.4 the left eigenstructure assignment is utilised to design robust PE
of a user defined order.

5.2 Robust fault detection via right eigenstructure assign-

ment for systems with unstable invariant zeros
Tan, Edwards & Kuang (2006b) extended the work of Chen & Patton (1999) providing a
right eigenstructure method for a sensor fault reconstruction. They demonstrated that
the invariant zeros of the transfer function between the disturbance and the output are
the unobservable modes of the robust fault detection filter. Therefore, filter stability can
only be ensured provided these zeros lie inside the unit circle. Tan, Edwards & Kuang
(2006b) used LMI to solve the filter equations. A continuation of they work has been
presented in (Tan, Edwards & Kuang 2006a), however the problem of using the right
eigenstructure assignment for the robust fault detection filter design in the presence of
unstable invariant zeros has not been resolved. The problem has been solved by Chen &
Speyer (2006a, 2006b, 2007) using a geometric approach. Chen & Speyer (2006b) used
the spectral theory to design a Beard-Jones fault isolation filter, whose applicability
has been extended to systems with unstable invariant zeros. Furthermore, a design of
the filter has been presented using eigenstructure assignment (Chen & Speyer 2006a)
and LMI (Chen & Nagarajaiah 2007).
This section extends the design method of a robust fault detection filter proposed
in (Chen & Patton 1999) to systems whose output response to disturbances contains
invariant zeros. Although the geometrical structure of the filter proposed here is similar
to that of Chen and colleagues (2006b, 2006a, 2007), the design procedure is simpler.
Similarities and differences between the robust fault detection filter presented in this
section and a fault isolation filter of Chen & Speyer (2006a) are discussed in Subsec-
tion 5.2.6.

5.2.1 Problem statement

It is assumed that a linear, dynamic, discrete-time, time-invariant system can be rep-
resented by the following equations, cf. (2.64):

x(t + 1) = Ax(t) + Bu(t) + Ed(t) + F µ(t)

(5.1)
y(t) = Cx(t) + Du(t)

108
 5. Robust fault detection via eigenstructure assignment

where x(t) ∈ Rn is the system state vector, u(t) ∈ Rp and y(t) ∈ Rm are, respectively,
the system input and output, d(t) ∈ Rq denotes a disturbance vector, whilst µ(t) ∈ Rr
is a fault signal. Matrices A, B, C, D, E, and F are constant and have appropriate
dimensions. It is assumed that (C, A) is an observable pair and the matrix E is of full
column rank.

Problem of unstable invariant zeros

Consider the robust fault detection filter described in Subsection 2.6.1 applied to the
system (5.1). Tan et al. (2006a) observed that the invariant zeros of the system
(A, Ei , C) are unobservable modes of the pair (C ′ , A′ ), cf. (2.69). As a result, if the
invariant zero is unstable, the pair (C ′ , A′ ) is not detectable. Consequently, in order to
ensure stability of the fault detection filter (2.54) in the case, when the triple (A, Ei , C)
has an unstable invariant zero, the design procedure needs to be altered, which is
proposed in Subsections 5.2.2, 5.2.3, and 5.2.4.
Lemma 5.1. Denote invariant zeros of (A, Ei , C) as z1 , z2 , ⋯, zqi . Then zeros of (A, Ei , C)
fulfil the following recursive set of equations:

⎡ ⎤⎡ ⎤
⎢ zj I − A −vj−1 ⎥ ⎢ vj ⎥
⎢ ⎥⎢ ⎥=0
⎢ ⎥⎢ ⎥
⎢ ⎥ ⎢ gj ⎥
(5.2)
⎣ ⎦⎣ ⎦
C 0

where v0 denotes Ei , vj , j = 1, ⋯, qi are vectors and gj , j = 1, ⋯, gi are scalar values.

Proof. Equation (5.2) can be reformulated as:

⎡ 0 ⎤
⎢ ⎥
⎢ ⎥
z1 −g2 0 ⋯
⎢ ⋯ 0 ⎥
⎢ ⎥
A [ v 1 v 2 ⋯ v qi ] − [ v 1 v 2 ⋯ v qi ]⎢ ⎥
0 z2 −g3
⎢ ⋱ ⋮ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋮ (5.3)
⎢ ⋯ zqi ⎥
⎣ 0 0 0 ⎦
+ Ei [ g 1 0 ⋯ 0 ] = 0

which is equivalent to (2.35).

Lemma 5.2. If the pair (A, C) is observable, scalars gj , j = 1, ⋯, qi in (5.2) are non-
zero.

Proof. If gj = 0 then (5.2) becomes:

⎡ ⎤
⎢ zj I − A ⎥
⎢ ⎥ vj = 0
⎢ ⎥
⎢ ⎥
(5.4)
⎣ ⎦
C

which means that zj is an unobservable mode of the system (A, Ei , C). However, the
pair (C, A) is observable, hence it does not have unobservable modes. Consequently,
gj ≠ 0.

109
 5. Robust fault detection via eigenstructure assignment

Rank condition

A certain case when the rank condition:

⎡ ⎤
⎢ Aλ ⎥
rank(CE) = rank(⎢
⎢
⎥)
⎥
⎢ CE ⎥
(5.5)
⎣ ⎦

is not fulfilled is considered here and a solution is proposed, which allows to relax the
strict rank condition and, consequently, apply the robust fault detection filter presented
in Subsection 2.6.1. Denote δi , i = 1, ⋯, q the lowest number for which:

CAδi Ei ≠ 0 (5.6)

Consider the situation when δi = 0 for i = 1, ⋯, q0 , where q0 < q, whilst δi ≠ 0 for

i = q0 + 1, ⋯, q. Then CEi = 0 for i = q0 + 1, ⋯, q, whilst the ith column of Aλ is:

(A − λi I)Ei (5.7)

The rank condition (5.5) is fulfilled if and only if the ith column of Aλ , i = q0 + 1, ⋯, q,
is zero. This occurs if and only if Ei is a right eigenvector of A corresponding to the
eigenvalue λi . This would, however, mean that Ei belongs to the unobservable subspace
of the pair (C, A). Thus, as it has been assumed that (C, A) is observable, i.e. (C, A)
does not have any unobservable subspace, Ei is not a right eigenvector of A and the
rank condition (5.5) is not fulfilled. In order to relax this condition it is proposed to
slightly alter the robust fault detection filter design presented in Subsection 2.6.1, by
replacing E with the following matrix:

e = [ e1 e2 ⋯ eq ] (5.8)

where:
ei = Aδi Ei (5.9)

From the definition of δi , cf. (5.6), it holds that Cei ≠ 0 for i = 1, ⋯, q. Therefore,
columns of Aλ do not require to be equal to zero.
Theorem 5.1. Invariant zeros of the triple (A, ei , C) are equal to the invariant zeros
of (A, Ei , C) plus δi zero-valued invariant zeros.

Proof. For j = 1, 2, ⋯, δi it holds, that

⎡ ⎤ ⎡ j−1 ⎤
⎢ −A −Aj Ei ⎥ ⎢ A Ei ⎥
⎢ ⎥⎢ ⎥=0
⎢ ⎥⎢ ⎥
⎢ C ⎥ ⎢ −1 ⎥
(5.10)
⎣ ⎦⎣ ⎦
0

Therefore, from Lemma 5.1 it follows, that invariant zeros of (A, ei , C) are z1 = z2 =
⋯ = zδi = 0, whilst the corresponding vectors v0 = Aδi Ei = ei , v1 = Aδi −1 Ei ,⋯, vδi = Ei ,
and g1 = g2 = ⋯ = gδi = −1.

110
 5. Robust fault detection via eigenstructure assignment

Consequently, replacing the matrix E with e in the filter described in Subsection 2.6.1
results in the pair (C ′ , A′ ) having zero-valued unobservable modes. It may be, however,
desired to set the eigenvalues of (A − KC) to different numbers that zero. In such a
situation the algorithms presented in the following subsections may be used.
Positive values of δi indicate a delay between the disturbance and the system output.
Thus, the system (5.1) can be reformulated as:

x(t + 1) = Ax(t) + Bu(t) + F µ(t) + ed∗ (t)

(5.11)
y(t) = Cx(t) + Du(t)

where elements of d∗ (t) are respective elements of d(t) delayed by δi , i.e.:

d∗ (t) = [ d1 (t − δ1 ) d2 (t − δ2 ) ⋯ dq (t − δq ) ]
T
(5.12)

5.2.2 Solution for q = 1 with a single invariant zero

In this subsection the algorithm of Chen & Patton (1999) is extended to the system,
where the invariant zero of (A, e, C) is unstable. For sake of simplicity it is assumed
that q = 1 and the triple (A, e, C) has only one invariant zero denoted as ze . Hence,
from the definition of an invariant zero, cf. (2.9), it follows that:

⎡ ⎤⎡ ⎤
⎢ ze I − A −e ⎥ ⎢ v ⎥
⎢ ⎥⎢ ⎥ = 0
⎢ ⎥⎢ ⎥
⎢ 0 ⎥⎢ g ⎥
(5.13)
⎣ ⎦⎣ ⎦
C

where v and g are the invariant zero state and input directions, respectively. Utilising
a similar solution to that in (Massoumnia 1986) a vector ē is created, such that:

ē = [ e v ] (5.14)

The aim of the scheme is to create such a filter that the state trajectory yielded by the
disturbance d(t) remains in the subspace Im{ē}, as opposed to the algorithm presented
in (Chen & Patton 1999), where the state trajectory of the disturbance d(t) remains in
the one-dimensional subspace Im{e}. In order for the solution to this problem to exist
the subspace Im{ē} must be (C, A)-invariant, i.e. there must exist such a gain matrix
K that Im{(A − KC)ē} ⊆ Im{ē}, see (Halmos 1958, Basile & Marro 2002). This means
that there exists such a matrix X, that, cf. (2.19):

(A − KC)ē = ēX (5.15)

Note that, if (A, e, C) has no invariant zeros, then v = ∅, and, consequently, ē = E and
X = λ1 , where λ1 is the desired eigenvalue of (A − KC) corresponding to the vector E.
The necessary and sufficient conditions for disturbance decoupling are:

1. The subspace Im{ē} is an invariant subspace of (A − KC)

111
 5. Robust fault detection via eigenstructure assignment

2. QCe = 0

From (5.15) it holds that the columns of ē are linear combinations of eigenvectors of
the matrix (A − KC):
ē = Ve Ψ (5.16)

where columns of Ve are the first two eigenvectors of (A − KC) and Ψ is an appropriate
matrix. Because (A − KC) is allocated distinct eigenvalues, rank{Ve } = 2, i.e. columns
of Ve are linearly independent. Also columns of ē are linearly independent (El-Ghezawi
et al. 1983). Consequently, matrix Ψ is of full rank. Furthermore:

(A − KC)Ve = Ve Λe (5.17)

where Λe is a diagonal matrix, whose diagonal elements are user defined eigenvalues
corresponding to the columns of Ve . By postmultiplying both sides of (5.17) by Ψ, the
following equation is obtained:

(A − KC)Ve Ψ = Ve Λe Ψ (5.18)

Incorporating (5.15) and (5.16) into (5.18):

Ve Λe Ψ = Ve ΨX (5.19)

This yields:
Λe Ψ = ΨX (5.20)

Therefore, recalling that Ψ is of full rank, the matrix X can be defined as:

X = Ψ−1 Λe Ψ (5.21)

where columns of Ψ−1 are right eigenvectors of X, whilst diagonal elements of Λe are
its corresponding eigenvalues. Consequently, it can be noted that the eigenvalues of X
are equal to the eigenvalues of (A − KC) corresponding to the columns of Ve , i.e. the
linear combinations of columns of ē.
From (5.15) it follows that:

KC ē = Aē − ēX (5.22)

Denote Aē − ēX as Ae . The necessary and sufficient conditions to assign all columns of
ē as linear combinations of the right eigenvectors of (A − KC) are:
⎡ ⎤
⎢ Ae ⎥
⎢
(i) rank(C ē) = rank(⎢ ⎥)
⎥
⎢ C ē ⎥
⎣ ⎦

112
 5. Robust fault detection via eigenstructure assignment

(ii) (C ′ , A′ ) is a detectable pair, where:

A′ = A − Ae (C ē) C
(5.23)
C ′ = (I − C ē(C ē) ) C

Theorem 5.2. Diagonal elements of Λe are unobservable modes of the pair (C ′ , A′ ).

Proof. Diagonal elements of Λe , denoted as λ1 and λ2 , correspond to the right eigen-

vectors of (A′ − K ′ C ′ ) denoted as w1 and w2 :

(A − Ae (Ce) C − K ′ (I − Ce(Ce) ) C) wi = λi wi (5.24)

which holds for any arbitrary K ′ , therefore if K ′ = 0:

(λi I − (A − Ae (Ce) C)) wi = (λi I − A′ )wi = 0 (5.25)

Consequently:
K ′ (I − Ce(Ce) ) Ce = K ′ C ′ wi = 0 (5.26)

which is valid for any K ′ , hence:

C ′ wi = 0 (5.27)

⎡ ⎤
As a result it holds that:
⎢ λi I − A′ ⎥
⎢ ⎥ wi = 0
⎢ ⎥
⎢ ⎥
(5.28)
⎣ ⎦
C′

which means that the diagonal elements of Λe are unobservable modes of the pair
(C ′ , A′ ) and only remaining n − 2 eigenvalues can be allocated by K ′ .

Calculation of the matrix X

From (5.13) it follows that:

Av = ze v − ge (5.29a)
Cv = 0 (5.29b)

Therefore, from (5.29b) it follows that:

(A − KC)v = Av (5.30)

Consider (5.15), then:

(A − KC) [ e v ] = [ e v ] X (5.31)

Denote elements of X as xij . Incorporating (5.30) into (5.31) it holds that:

[ (A − KC)e Av ] = [ x11 e + x21 v x12 e + x22 v ] (5.32)

113
 5. Robust fault detection via eigenstructure assignment

Consequently, from (5.29a) it can be deduced that x12 = −g and x22 = ze . Knowing that
X has the same eigenvalues as Λe , x11 and x21 are calculated as:

x11 = λ1 + λ2 − ze (5.33a)
(ze − λ1 )(ze − λ2 )
x21 = − (5.33b)
g

Rank condition

Recall the rank condition (i) in Subsection 5.2.1 and (5.29b). It can be deduced that:

rank(C ē) = rank([ Ce Cv ]) = rank(Ce) (5.34)

and:
Ae = [ Ae − x11 e − x21 v Av − ze v + ge ] (5.35)

From (5.29a), it holds that the second column of Ae is equal to zero. Therefore, using
the notation:
A∗e = Ae − x11 e − x21 v (5.36)

⎡ ⎤ ⎡ ∗ ⎤
it holds that:
⎢ Ae ⎥ ⎢ ⎥
rank(⎢
⎢
⎥) = rank(⎢ Ae ⎥)
⎥ ⎢ ⎥
⎢ C ē ⎥ ⎢ Ce ⎥
(5.37)
⎣ ⎦ ⎣ ⎦
Hence, the assignability condition can be reformulated as:
⎡ ∗ ⎤
⎢ Ae ⎥
(i) rank(Ce) = rank(⎢
⎢
⎥)
⎥
⎢ Ce ⎥
⎣ ⎦
(ii) (C ′ , A′ ) is a detectable pair, where:

A′ = A − A∗e (Ce) C
(5.38)
C ′ = (I − Ce(Ce) ) C

The algorithm for the design of a robust fault detection filter (RFDF) using right
eigenstructure assignment is summarised below.

Algorithm 5.1 (RFDF via right eigenstructure assignment, q = 1, q1 = 1).

1. Obtain disturbance direction matrix e using (5.6), (5.8) and (5.9)

2. Calculate Q such that QCe = 0

3. Select eigenvalues λ1 and λ2

114
 5. Robust fault detection via eigenstructure assignment

4. Calculate invariant zero state and input directions v and g from

⎡ ⎤⎡ ⎤
⎢ ze I − A −e ⎥ ⎢ v ⎥
⎢ ⎥⎢ ⎥ = 0
⎢ ⎥⎢ ⎥
⎢ 0 ⎥⎢ g ⎥
(5.39a)
⎣ ⎦⎣ ⎦
C

5. Calculate coefficients x11 and x21 via:

x11 = λ1 + λ2 − ze (5.39b)
(ze − λ1 )(ze − λ2 )
x21 = − (5.39c)
g

6. Obtain matrix Ae as:

A∗e = (A − x11 I)e − x21 v (5.39d)

7. Obtain:

A′ = A − A∗e (Ce) C (5.39e)

C ′ = (I − Ce(Ce) ) C (5.39f)

8. Using any eigenstructure assignment method allocate remaining n − 2 eigen-

values of (A′ − K ′ C ′ )

9. Calculate the gain matrix K:

K = A∗e (Ce) + K ′ (I − Ce(Ce) ) (5.39g)

5.2.3 Solution for q = 1 with multiple invariant zeros

Consider the system (5.1) where q = 1 and the triple (A, e, C) has q1 (q1 < n) invariant
zeros denoted as z1 , ⋯, zq1 . One can assign a vector vi and a scalar gi to each invariant

⎡ ⎤⎡ ⎤
zero zi such that:
⎢ zi I − A −vi−1 ⎥ ⎢ vi ⎥
⎢ ⎥⎢ ⎥=0
⎢ ⎥⎢ ⎥
⎢ ⎥ ⎢ gi ⎥
(5.40)
⎣ ⎦⎣ ⎦
C 0

where v0 refers to the vector e. It follows from Lemma 5.2 that gi are non-zero, hence,
vectors vi , i = 1, 2, ⋯, q1 can be scaled such that gi = −1 for i = 1, 2, ⋯, q1 . The aim of the
algorithm is to force the state trajectory governed by the disturbance d(t) to remain
within the subspace Im{ē} defined as:

ē = [ e v1 ⋯ vq1 ] (5.41)

115
 5. Robust fault detection via eigenstructure assignment

which requires ē to be a (C, A)-invariant subspace, i.e. there exist such a gain matrix
K that, cf. (5.15):
KC ē = Aē − ēX (5.42)

Due to the fact that Cvi = 0 for i = 1, ⋯, q1 , all columns of Ae = Aē − ēX, except of the
first one, must be equal to zero for the solution of (5.42) to exist. Denote the elements
of X as xij , then the ith column of Ae is given by:

Avi−1 − (x1i v0 + x2i v1 + ⋯ + xq1 +1 vq1 ) = 0 (5.43)

From (5.40) it follows that (recall that gi = −1 for i = 1, 2, ⋯, q1 ):

Avi−1 = zi−1 vi−1 + vi−2 (5.44)

Incorporating (5.44) into (5.43) the following equation is obtained:

x1i v0 + x2i v1 + ⋯ + (xi−1,i − 1)vi−2 + (xi,i − zi−1 )vi−1 + ⋯ + xq1 +1 vq1 = 0 (5.45)

Due to the fact that v1 , ⋯, vq1 are linearly independent:

xi−1,i = 1 for i = 2, ⋯, q1 + 1 (5.46a)

xi,i = zi−1 for i = 2, ⋯, q1 + 1 (5.46b)
xj,i = 0 for j ≠ i and j ≠ i − 1 (5.46c)

Consequently the matrix X is given by:

⎡ x 0 ⎤
⎢ ⎥
⎢ ⎥
1 0 0 ⋯ 0
⎢ x 0 ⎥
11

⎢ ⎥
⎢ ⎥
z1 1 0 ⋯ 0
⎢ x 0 ⎥
21

⎢ ⎥
X =⎢ ⎥
31 0 z2 1 ⋯ 0
⎢ ⋮ ⎥
(5.47)
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋮ ⋮ ⋱ ⋮
⎢ xq1 1,1 0 0 1 ⎥
⎢ 0 ⋯ zq1 −1 ⎥
⎢ ⎥
⎢ xq1 +1,1 0 0 z q1 ⎥
⎣ 0 ⋯ 0 ⎦

The first column of X is chosen such that the eigenvalues of X are equal to the desired
eigenvalues of (A − KC) corresponding to the linear combinations of columns of ē,
cf. (5.20). Consequently, (for derivation details see Appendix A) the first column of X
given by:
⎡ x ⎤
⎢ ⎥
⎢ ⎥
⎢ x ⎥
11

⎢ ⎥ ˜−1 ˜
⎢ ⎥ = Ã B̃
12
⎢ ⎥
(5.48)
⎢ ⎥
⎢ ⎥
⋮
⎢ xq1 +1,1 ⎥
⎣ ⎦

116
 5. Robust fault detection via eigenstructure assignment

where an element of Ø ∈ R(q1 +1)×(q1 +1) , denoted as Øj,k is:

q
Øj,k = (−1)k−1 ∏(zl − λj )
i
(5.49)
l=k

˜ ∈ Rq1 +1 , denoted as B̃
whilst the j th element of the vector B̃ ˜ is:
j

q
˜ = λ i (z − λ )
B̃ j j∏ l j (5.50)
l=1

Algorithm 5.2 (RFDF via right eigenstructure assignment, q = 1, q1 ≥ 1).

1. Obtain disturbance direction matrix e using (5.6), (5.8) and (5.9)

2. Obtain Q such that QCe = 0

3. Denote e as v0 and obtain invariant zeros of the pair (A, e, C) and corre-
sponding vectors vi , for i = 1, ⋯, q1

⎡ ⎤⎡ ⎤
⎢ zi I − A −vi−1 ⎥ ⎢ vi ⎥
⎢ ⎥⎢ ⎥=0
⎢ ⎥⎢ ⎥
⎢ ⎥ ⎢ −1 ⎥
(5.51a)
⎣ ⎦⎣ ⎦
C 0

4. Select eigenvalues λ1 , ⋯, λq1 +1

5. Calculate coefficients x11 , ⋯, xq1 +1,1 using (5.48), (5.49), and (5.50)

6. Obtain matrix A∗e as:

A∗e = (A − x11 I)e − x21 v1 − x31 v2 − ⋯ − xq1 +1,1 vq1 (5.51b)

7. Obtain:

A′ = A − A∗e (Ce) C (5.51c)

C ′ = (I − Ce(Ce) ) C (5.51d)

8. Allocate remaining n − q1 − 1 eigenvalues of (A′ − K ′ C ′ )

9. Calculate K
K = A∗e (Ce) + K ′ (I − Ce(Ce) ) (5.51e)

117
 5. Robust fault detection via eigenstructure assignment

5.2.4 General solution for q ≥ 1

In this section a general solution for the robust fault detection filter when q ≥ 1 is
presented. It is assumed that the triple (A, e, C) has invariant zeros. The invariant
zeros of (A, ei , C), where ei refers to the ith column of e, are denoted as z1 , ⋯, zqi . It
(i) (i)

is assumed that the invariant zeros fulfil the condition:

q
Ω = ⊎ Ωi (5.52)
i=1

where Ω is the set of all invariant zeros of (A, e, C), whilst Ωi denotes the set of the
invariant zeros of (A, ei , C), see (Massoumnia 1986). The aim of the algorithm is to
ensure that the state trajectory driven by di (t) remains in the subspace Im{ēi }, where:

ēi = [ ei v1(i) ⋯ vq(i)

i
] (5.53)

Therefore, the necessary and sufficient conditions for the robust fault detection are:

1. For each column of the matrix e it holds that Im{e¯i } is an invariant subspace of
(A − KC)

2. QCe = 0

Analogously to the case where e is a column vector, cf. Subsection 5.2.2, the matrix
A∗e is built, such that:
A∗e = [ A∗(1)
e
∗(2)
Ae
∗(q)
⋯ Ae ] (5.54)

where:
= (A − x11 I)ei − x21 v1 − x31 v2 − ⋯ − xqi +1,1 vq(i)
(i) (i) (i) (i) (i) (i)
A∗(i)
e i
(5.55)

The matrices A′ and C ′ are built as in (5.51c) and (5.51d). Note that if the (A, ei , C)
has no invariant zeros then it holds that δi = 0 and ei = Ei , see Lemma 5.2, and:

A∗(i)
e = (A − λi I)ei (5.56)

Hence, if the system has no invariant zeros and δi = 0 for i = 1, ⋯, q the algorithm
presented here is equivalent to that of Chen & Patton (1999). The necessary conditions
for the solution of the robust fault detection filter to exist are:
⎡ ∗ ⎤
⎢ Ae ⎥
(i) rank(Ce) = rank(⎢ ⎢
⎥)
⎥
⎢ Ce ⎥
⎣ ⎦
(ii) (C ′ , A′ ) is a detectable pair, where:

A′ = A − A∗e (Ce) C
(5.57)
C ′ = (I − Ce(Ce) ) C

118
 5. Robust fault detection via eigenstructure assignment

Similarly as in the previous case the matrix X (i) is given by:

⎡ x(i) 0 ⎤
⎢ ⎥
⎢ ⎥
1 0 0 ⋯ 0
⎢ x(i) 0 ⎥
11
⎢ ⎥
(i)
⎢ ⎥
z1 1 0 ⋯ 0
⎢ x (i) 0 ⎥
21
⎢ ⎥
(i)
X (i) = ⎢ 31 ⎥
0 z2 1 ⋯ 0
⎢ ⋮ ⎥
(5.58)
⎢ ⎥
⎢ (i) ⎥
⋮ ⋮ ⋮ ⋮ ⋱ ⋮
⎢ x 1 ⎥
⎢ q1 1,1 ⎥
(i)
0 0 0 ⋯ zq1 −1
⎢ (i) (i) ⎥
⎢ x z q1 ⎥
⎣ q1 +1,1 0 0 0 ⋯ 0 ⎦

and its first column is calculated via:

⎡ x(i) ⎤
⎢ ⎥
⎢ ⎥
⎢ x(i) ⎥
11
⎢ ⎥
⎢ ⎥ = (Ø(i) )−1 B̃
˜ (i)
⎢ ⎥
12
(5.59)
⎢ ⎥
⎢ (i) ⎥
⋮
⎢ x ⎥
⎣ q1 +1,1 ⎦

where an element of Ø(i) ∈ R(qi +1)×(qi +1) , denoted as Øj,k is:

(i)

q
Øj,k = (−1)k−1 ∏(zl − λj )
i
(i) (i) (i)
(5.60)
l=k

˜ ∈ Rq1 +1 , denoted as B̃
whilst the j th element of the vector B̃ ˜ is:
j

q
˜ (i) = λ(i) i (z (i) − λ(i) )
B̃ j j ∏ l j (5.61)
l=1

The generalised form of the algorithm for disturbance decoupled fault detection
filter is given below.

Algorithm 5.3 (RFDF via right eigenstructure assignment, q ≥ 1, qi ≥ 0).

1. Calculate disturbance direction matrix e using (5.6), (5.8) and (5.9)

2. Obtain Q such that QCe = 0

3. For each column of e obtain invariant zeros of the triple (A, ei , C), denoted
(i) (i)
as zj , and corresponding vectors vj , for j = 1, ⋯, qi

⎡ (i) ⎤ ⎡ (i) ⎤
⎢ zj I − A −vj−1 ⎥ ⎢ vj ⎥
⎢ ⎥⎢ ⎥=0
(i)

⎢ ⎥⎢ ⎥
⎢ ⎥ ⎢ −1 ⎥
(5.62a)
⎣ C 0 ⎦⎣ ⎦
(i)
where v0 denotes ei .

119
 5. Robust fault detection via eigenstructure assignment

(i) (i) (i)

4. Select eigenvalues λ1 , ⋯, λqi +1 corresponding to linear combinations of vj ,
j = 1, 2, ⋯, qi + 1
(i) (i)
5. Calculate coefficients x11 , ⋯, xqi +1,1 using (5.59–5.61)

6. Obtain matrix A∗e as:

A∗e = [ A∗(1)
e
∗(2)
Ae
∗(q)
⋯ Ae ] (5.62b)

where

= (A − x11 I)ei − x21 v1 − x31 v2 − ⋯ − xqi +1,1 vk

(i) (i) (i) (i) (i) (i) (i)
A∗(i)
e (5.62c)

7. Obtain:

A′ = A − A∗e (Ce) C (5.62d)

C = (I − Ce(Ce) ) C
′
(5.62e)

8. Allocate remaining eigenvalues of (A′ − K ′ C ′ )

9. Calculate K
K = A∗e (Ce) + K ′ (I − Ce(Ce) ) (5.62f)

5.2.5 Consideration of residual response to fault

In this subsection some remarks considering the residual response to a fault and its
dependency on the choice of the gain matrix K are discussed.

Zero-pole cancellation in the residual response to fault

Remark 5.1. Unobservable modes of (C ′ , A′ ) are unobservable modes of (QC, (A −

KC)).

Demonstration. Consider unobservable modes of (C ′ , A′ ), i.e. the eigenvalues of

(A − KC) corresponding to linear combinations of ei . Denote the eigenvectors of (A −
KC) corresponding to the unobservable modes of (C ′ , A′ ) as wj , i = 1, ⋯, q; j = 1, ⋯, qi .
(i)

]}. It is known that:

(i)
Recall that wj ∈ Im{[ ei v1(i) ⋯ vq(i)
i

(i)
Cvj = 0 (5.63a)
QCei = 0 (5.63b)

120
 5. Robust fault detection via eigenstructure assignment

Therefore, it holds that:

(i)
QCwj = 0 (5.64a)
(Iλj − A + KC) wj = 0
(i) (i)
(5.64b)

is the eigenvalue of (A − KC) corresponding to the eigenvector λj . Conse-

(i) (i)
where λj
quently, the unobservable modes of (C ′ , A′ ) are the unobservable modes of (QC, (A −
KC)).

This results in a zero-pole cancellation of the unobservable modes of (C ′ , A′ ) in the

transfer function of (A − KC, F, QC). Consequently, the observable modes of (A −
KC, F, QC) are only those eigenvalues of (A − KC) which are assigned by the choice
of the matrix K ′ . This information may be useful for designing the residual response
to faults.
Note that the necessary condition for the fault to be detected by the filter is:

QC(zI − A + KC)−1 F ≠ 0 (5.65)

Without loss of generality assume that F is a column vector. Using the notation
ē = [ ē1 ē2 ⋯ ēq ], the matrix F can be expressed as a sum of its orthogonal
projections on Im{ē} and the orthogonal completion of Im{ē}

F = F (ē) + F (ē– ) (5.66)

where F (ē) is an orthogonal projection of F on Im{ē}, whereas F (ē– ) denotes an

orthogonal projection of F on the orthogonal completion on Im{ē}. Due to the fact
that F (ē) belongs to the unobservable subspace of the fault detection filter, the fault
to residual transfer function, denoted as Gf r (z), is given by:

Gf r (z) = QC(zI − A + KC)−1 F (ē– ) (5.67)

This means that the necessary condition for the robust fault detection filter to exist
is that the dimension of the unobservable subspace of (QC, A − KC) is lower that n
(otherwise no fault can be detected as the whole state space is unobservable for the
fault detection filter).

Invariant zeros in the residual response to fault

Invariant zeros shape the response of the residual to a fault. In some situations, e.g.
fault identification, it may be desirable to influence not only its poles, but also zeros.

Remark 5.2. If rank(Q) = n−rank(Ce), then the selection of eigenvalues of (A−KC)

has no influence on the invariant zeros of the residual response to a fault.

121
 5. Robust fault detection via eigenstructure assignment

Demonstration. Without loss of generality it is assumed that dim{F } = 1. Then the

invariant zero of the residual response to a fault, denoted zf , fulfils the condition:

⎡ ′ ⎤⎡ ⎤
⎢ A − K ′ C ′ − zf I F ⎥ ⎢ v f ⎥
⎢ ⎥⎢ ⎥=0
⎢ ⎥⎢ ⎥
⎢ 0 ⎥ ⎢ gf ⎥
(5.68)
⎣ ⎦⎣ ⎦
QC

where vf and gf are the invariant zero state and input directions, respectively. Note
that Im{vf } ⊂ Ker{QC}, i.e. the state direction vf belongs to the right nullspace of QC.
The matrix Q which fulfils the condition QCe = 0 can be defined as (Basilevsky 1983):

Q = Q0 (I − Ce(Ce) ) (5.69)

where Q0 is an arbitrary matrix. Note that rows of Q are linear combinations of rows
of (I − Ce(Ce) ). Therefore, if rank(Q0 ) = rank (I − Ce(Ce) ) = n − rank(Ce), then
rank(Q) = n − rank(Ce). Hence, it follows, that if rank(Q0 ) = rank (I − Ce(Ce) ),
then the subspace spanned by the rows of Q is the subspace spanned by the rows of
(I − Ce(Ce) ). Furthermore, it holds that:

QC = Q0 (I − Ce(Ce) ) C = Q0 C ′ (5.70)

Hence, Ker{QC} = Ker{Q0 C ′ }. This means that, if rank(Q) = n − rank(Ce) then

Ker{QC} = Ker{Q0 C ′ } = Ker{C ′ }. Recall that vf ⊂ Ker{QC}, then, if rank(Q) =
n − rank(Ce), it holds that:
QCvf = C ′ vf = 0 (5.71)

This shows that the invariant zero of the residual response to fault, defined by (5.68),
does not depend on choice of K ′ , i.e. zf does not depend on the choice of the eigenvalues
of (A − KC), which are not corresponding to the linear combinations of ēi , i = 1, ⋯, q.
As a result (5.68) can be rewritten as:

A′ vf − zf vf + F gf = 0 (5.72a)
QCvf = 0 (5.72b)

Now it will be demonstrated that the invariant zeros of the residual response to
a fault do not depend on the choice of eigenvalues of (A − KC) corresponding to the
linear combinations of ēi , i = 1, ⋯, q. Consider the following change of basis using the
following orthonormal matrix:
T = [ T 1 T2 ] (5.73)

where:
Im{T1 } = Im{[ ē1 ē2 ⋯ ēq ]} (5.74)

122
 5. Robust fault detection via eigenstructure assignment

and Im{T2 } is an orthogonal completion of Im{T1 }. Equation (5.72) is reformulated

using the similarity transformation T :

T T A′ T T T vf − zf T T vf + gf T T F = 0 (5.75a)
QCT T T vf = 0 (5.75b)

which furthermore can be rewritten as:

⎡ ⎤⎡ ⎤ ⎡ T ⎤ ⎡ T ⎤
⎢ A1 A2 ⎥ ⎢ T1T vf ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥⎢ ⎥ − z f ⎢ T1 v f ⎥ + g f ⎢ T1 F ⎥=0
⎢ ⎥⎢ ⎥ ⎢ T ⎥ ⎢ T ⎥
⎢ 0 A3 ⎥ ⎢ T2T vf ⎥ ⎢ T2 v f ⎥ ⎢ T2 F ⎥
(5.76a)
⎣ ⎦⎣ ⎦ ⎣ ⎦ ⎣ ⎦
⎡ T ⎤
⎢ T1 v f ⎥
[ 0 QCT2 ] ⎢ ⎢ T
⎥=0
⎥
⎢ T2 v f ⎥
(5.76b)
⎣ ⎦

where A1 , A2 , and A3 are the appropriate submatrices of T T A′ T . The first element

in the left hand side matrix of (5.76b) is equal zero because QCe = 0. Using the
notation vf′ = T2T vf , if the invariant zero of the residual response to a fault exists, then
it conforms the following equation:

A3 vf′ − zf vf′ + F gf = 0 (5.77a)

QCvf′ = 0 (5.77b)

Knowing vf′ one can calculate vf which fulfils vf′ = T2T vf by solving dim{T2 } equations
with n unknowns. Hence the obtained solution has n−dim{T2 } parameters. The second
part of (5.76), i.e.

(A1 T1T + A2 T2T − zf T1T + gf T1T ) vf = 0 (5.78a)

consists of n − dim{T2 } equations, from which remaining parameters can be found and
the vector vf calculated. Therefore, the existence of zf depends on A3 = T2T A′ T2 .
Recall (5.62b) and (5.62c):

A∗e = [ Ae1 Ae2 ⋯ Aeq ] + [ ∑qi=0

1 (1)
vi q2
∑i=0
(2)
vi
qq
⋯ ∑i=0 vi ]
(q)
(5.79)
= Ae + Ω

where:
Ω = [ ∑qi=0
1 (1)
vi q2
∑i=0
(2)
vi
qq
⋯ ∑i=0 vi ]
(q)
(5.80)

Recall that:
A′ = A − Ae(Ce) C + Ω(Ce) C (5.81)

Due to the fact that T2T Ω = 0:

A3 = T2T A′ T2 = T2T (A − Ae(Ce) C)T2 (5.82)

123
 5. Robust fault detection via eigenstructure assignment

Therefore, the matrix A3 and hence the invariant zeros of the residual response to a
fault do not depend on the choice of the eigenvalues of (A − KC).

5.2.6 Differences and similarities with fault isolation filter of Chen

and Speyer (2006a)
Although Algorithm 5.3 is designed for a robust fault detection, whilst Algorithm pre-
sented in Subsection 2.7.1, cf. (Chen & Speyer 2006a), is for a fault isolation, their
eigenstructures are the same, i.e. the eigenstructure of (A − KC, E, C) using Algo-
rithm 5.3 and eigenstructure of (A − KC, F, C) using the algorithm of Chen & Speyer
(2006b) are the same. The idea of both schemes is to find such a gain matrix K that
the following conditions are fulfilled:

(i) Eigenvalues of (A−KC) can be arbitrarily chosen (with constraint to no repeated

eigenvalues and conjugate symmetry)

(ii) Eigenvectors of (A − KC) are linear combinations of columns of e (f ) and their

invariant zeros state directions.

Note that the condition (ii) in (Chen & Speyer 2006a) has been specified as:

(i) (i) (i)

wj = Θ̄(i) β̄j + fi = Θ(i) βj (5.83)

is an eigenvector of (A − KC) and β̄ (i) is an appropriate coefficient vector.

(i)
where wj
Columns of Θ̄(i) span the following subspace:

Im{Θ̄(i) } = Im{[ Fi AFi ⋯ Aδi −1 Fi ]} ⊕ Vi (5.84)

where Vi is the subspace spanned by invariant zero state directions of (A, Fi , C).
Chen & Speyer (2006a) explicitly indicate that each eigenvector of (A − KC) corre-
sponding to fi contains the vector fi . Although it is not explicitly said Algorithm 5.3
is characterised by the same property.

Lemma 5.3. Each eigenvector of (A − KC) obtained using Algorithm 5.3 correspond-
ing to linear combination of e¯i contains ei .

Proof. Denote an eigenvector of (A − KC) corresponding to a linear combination of

columns of e¯i as:
(i) (i) (i) (i) (i)
wj = α0 ei + α1 v1 + α2 v2 + ⋯ + αq(i)
i
vq(i)
i
(5.85)

Consider the situation when α0 = 0, i.e. the j th eigenvector of (A−KC) corresponding

(i)

′(i) (i)
to linear combination of ē, denoted as wj , is a linear combination of vectors vk , k =
1, 2, ⋯, qi but not ei :
(i) (i) (i)
= α1 v1 + α2 v2 + ⋯ + αq(i) vq(i)
′(i)
wj i i
(5.86)

124
 5. Robust fault detection via eigenstructure assignment

Then it holds that:

(A − KC)wj
′(i) (i)
= λj w′(i) (5.87)
′(i) (i) ′(i)
As wj is a linear combination of vk , k = 1, ⋯, qi , it holds that Cwj = 0, cf. (5.40).
Consequently:
′(i) (i) ′(i)
Awj = λj w j (5.88)
′(i) ′(i)
i.e. wj a right eigenvector of A corresponding to eigenvalue λj . Then, it holds that:

(i) (i) (i) (i) (i) (i) (i) (i) (i)

α1 Av1 + α2 Av2 + ⋯ + αq(i)
i
Avq(i)
i
= λj v1 + λj v2 + ⋯ + λj vq(i)
i
(5.89)

Incorporating (5.40) into (5.89):

− α1 g1 ei + (α1 z1 − α2 g2 − λj ) v1 + (α2 z2 − α3 g3 − λj ) v2 + ⋯+
(i) (i) (i) (i) (i) (i) (i) (i) (i) (i) (i) (i)

(5.90)
(αqi −1 zqi −1 − αq(i) gqi − λj ) vqi −1 + (αq(i) − λj ) vq(i)
(i) (i) (i) (i) (i)
i i
zq(i)
i i
=0

Because columns of ēi are linearly independent, the above equation holds if and only
if:

(i) gi = 0

(or)
(i) (i) (i)
(ii) α1 = α2 = ⋯ = αqi = 0

Assumption (i) does not hold as the system is observable, see Lemma 5.2. Assumption
(ii) would mean that the eigenvector of (A−KC) corresponding to a linear combination
of columns of ēi is equal zero. Consequently, eigenvectors of (A − KC) corresponding
to linear combination of e¯i must contain ei . Due to the fact that eigenvectors can be
(i)
arbitrarily scaled, the coefficient α0 can be set to unity and the rest of the coefficients
can be scaled accordingly.

The solution to Algorithm 5.3 is such a matrix K that:

KCe = A∗e (5.91)

whilst the solution to the algorithm of Chen & Speyer (2006a) is:

KCf = Aw (5.92)

where

Aw = [ (A − λ(1) (1)
1 I)w1 (A − λ1 I)w1
(2) (2)
⋯ (A − λ1 I)w1 ]
(q) (q)
(5.93)

and w1 , i = 1, 2, ⋯, q, are eigenvectors of (A − KC) corresponding to λj .

(i) (i)

125
 5. Robust fault detection via eigenstructure assignment

Lemma 5.4. Matrices A∗e and Aw are equal.

Proof. See Appendix B.

The main difference between both algorithms is the calculation of the coefficients
needed to obtain columns of A∗e (Aw ). The algorithm presented in Subsection 2.7.1
requires to calculate qi (qi + 1) of β̄j coefficients in order to calculate one column of
Aw . This is done by a pseudoinverse of a matrix of the dimension nqi × qi (qi + 1).
Furthermore, the β̄j coefficients are linearly dependent, cf. (B.9), and not all of them
are needed to compute a column of Aw . On the other hand, Algorithm 5.3 requires qi +1
coefficients to obtain any column of A∗e , which are calculated by solving a set of qi + 1
linear equations, which requires an inverse of a matrix of the dimension (qi +1)×(qi +1),
which is computationally less demanding compared to the algorithm of Chen & Speyer
(2006a).

5.2.7 Tutorial examples

Example 5.1. q = 1, single invariant zero
Consider the system (5.1), whose A, C, D, E, and F matrices are given by:

⎡ 0 ⎤ ⎡ −0.8000 ⎤ ⎡ 1 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
1 0 0
⎢ 0 ⎥ ⎢ 1.4000 ⎥ ⎢ 1.1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
A=⎢ ⎥ E=⎢ ⎥ F =⎢ ⎥
0 1 0
⎢ 0 ⎥ ⎢ 1.2000 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ (5.94)
0 0 1
⎢ −0.1155 −0.7985 −2.06 −2.35 ⎥ ⎢ 3.7725 ⎥ ⎢ 0 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
⎡ ⎤
⎢ −0.8165 0.5266 −0.2367 0 ⎥
⎢
C =⎢ ⎥
⎥
⎢ −0.4082 −0.2367 0.8816 0 ⎥
⎣ ⎦

The triple (A, E, C) has one invariant zero at ze = 1.2, whose input direction is g = 1
and the zero state direction is given by:

v=[ 1 2 1 0 ]
T
(5.95)

Also CE ≠ 0, thus δ1 = 0 and e = E. The aim of the algorithm is to limit the state
trajectory governed by d(t) to the subspace Im{[ E v ]}. Eigenvalues of (A − KC)
corresponding to linear combinations of E and v are selected to be 0.45 and 0.65. This
corresponds to x11 = −0.1, cf. (5.39b), and x21 = 0.4125, cf. (5.39c). Consequently, the
matrix A∗e is calculated as, cf. (5.39d):

⎡ 0.9075 ⎤
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎢ ⎥
A∗e = ⎢ ⎥
0.5150
⎢ 3.4800 ⎥
(5.96)
⎢ ⎥
⎢ ⎥
⎢ −11.9856 ⎥
⎣ ⎦

126
 5. Robust fault detection via eigenstructure assignment

Matrices A′ and C ′ are given by, cf. (5.39e) and (5.39f):

⎡ 0.5186 ⎤
⎢ ⎥
⎢ ⎥
0.8704 −0.2593 0
⎢ 0.2943 −0.0736 ⎥
′ ⎢ ⎥
A =⎢ ⎥
0.8529 0
⎢ 1.9886 −0.4971 −0.9943 ⎥
⎢ ⎥
⎢ ⎥
1
⎢ −6.9644 ⎥
(5.97)
⎣ 0.9137 1.3645 −2.35 ⎦
⎡ ⎤
⎢ ⎥
C′ = ⎢ ⎥
−0.1843 0.3685 −0.5528 0
⎢ ⎥
⎢ 0.1936 −0.3872 0 ⎥
⎣ ⎦
0.5807

⎡ ∗ ⎤
It is noted that:
⎢ Ae ⎥
rank(CE) = rank(⎢
⎢
⎥) = 1
⎥
⎢ CE ⎥
(5.98)
⎣ ⎦
and (C ′ , A′ ) is a detectable pair, i.e. its unobservable modes, 0.45 and 0.65, lie within
the unit circle. Therefore, the solution for the stable filter design exists. The remaining
eigenvalues of (A − KC) are chosen to be 0.35 and 0.4 and consequently the matrix K ′
is given by:
⎡ −0.5936 4.0532 ⎤
⎢ ⎥
⎢ ⎥
⎢ 0.7442 2.6206 ⎥
′ ⎢ ⎥
K =⎢ ⎥
⎢ 0.6739 −7.8011 ⎥
(5.99)
⎢ ⎥
⎢ ⎥
⎢ −0.9526 2.0485 ⎥
⎣ ⎦
Finally, the gain matrix K is obtained as:

⎡ −1.8760 2.8324 ⎤
⎢ ⎥
⎢ ⎥
⎢ −0.7107 ⎥
⎢ ⎥
K =⎢ ⎥
1.2356
⎢ 5.8663 −2.8584 ⎥
(5.100)
⎢ ⎥
⎢ ⎥
⎢ −7.1590 −3.8596 ⎥
⎣ ⎦

Note that the fault distribution matrix F can be represented as a sum of:

⎡ 0.3889 ⎤ ⎡ 0.6111 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.0528 ⎥ ⎢ 1.0472 ⎥
⎢ ⎥ ⎢ ⎥
F (Ē ) = ⎢ ⎥ and F (Ē) = ⎢ ⎥
⎢ −0.4944 ⎥ ⎢ 0.4944 ⎥
–
(5.101)
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.2201 ⎥ ⎢ −0.2201 ⎥
⎣ ⎦ ⎣ ⎦

where F (Ē) ∈ Im{Ē}, whilst the vector F (Ē – ) is orthogonal to the subspace Im{Ē}.
Furthermore, the generalised angle between and F and the subspace Im{Ē} is given
by:
F T F (Ē) 180o
arccos ( ) = 26.7o
∣∣F ∣∣2 ∣∣F (Ē)∣∣2
(5.102)
π
Consequently, as the fault input direction F ∉ Im{Ē}, the fault occurrence can be
detected by the robust fault detection filter. The simulation results are presented in
Fig. 5.1. As expected the fault detection filter is insensitive to disturbances, whereas

127
 5. Robust fault detection via eigenstructure assignment

Disturbance
0.2

0.1
d(t)

−0.1

−0.2
0 10 20 30 40 50 60 70 80 90 100

Fault detection
1.5
µ(t), r(t)

0.5

0
µ(t)
r(t)
−0.5
0 10 20 30 40 50 60 70 80 90 100
Time [samples]

Figure 5.1: Robust fault detection, q = 1, the triple (A, E, C) has single invariant zero.
Upper subfigure presents disturbances, whilst lower subfigure demon-
strates robust fault detection process. It can be noted that residual,
r(t), is insensitive to disturbances.

it is sensitive to fault.

Example 5.2. q = 1, multiple invariant zeros

In this example q = 1 and the triple (A, E, C) has two invariant zeros. Matrices of the
state-space system (5.1) are given by:

⎡ 0 ⎤ ⎡ −2.4400 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
1 0 0 0
⎢ 0 ⎥ ⎢ 0.6200 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
0 1 0 0
⎢
A=⎢ 0 ⎥ E = ⎢ 1.5600 ⎥
⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
0 0 1 0
⎢ 0 ⎥ ⎢ −3.4816 ⎥
⎢ 0 0 0 1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ (5.103)
⎢ −0.0751 −0.6345 −2.1375 −3.5875 −3 ⎥ ⎢ 18.3643 ⎥
⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤T
⎢ 0.0948 −0.4811 0.8675 0 −0.0837 ⎥ ⎢ 1 −1 ⎥
⎢
C =⎢ ⎥ F =⎢
1 1 0.3 ⎥
⎥ ⎢ ⎥
⎢ −0 ⎥ ⎢ 2 ⎥
⎣ ⎦ ⎣ ⎦
0 0 1 0 3.4 2 0 0

The triple (A, E, C) has two invariant zeros, namely z1 = 1.3 and z2 = 1.2, and:

⎡ 0.8000 ⎤ ⎡ 1 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ −1.4000 ⎥ ⎢ 2 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
v1 = ⎢ −1.2000 ⎥ v2 = ⎢
⎢ ⎥
⎢ 1 ⎥
⎥
⎢ ⎥ ⎢ ⎥
(5.104)
⎢ 0.0000 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ −3.4816 ⎥ ⎢ 0 ⎥
⎣ ⎦ ⎣ ⎦

128
 5. Robust fault detection via eigenstructure assignment

Poles of (A − KC) corresponding to linear combinations of E, v1 , and v2 are selected

to be 0.35, 0.3, and 0.25. Consequently, matrices Ø and B̃
˜ are, cf. (5.49) and (5.50):

⎡ 0.9975 −0.95 1 ⎤ ⎡ 0.2494 ⎤

⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
˜ ⎢
à = ⎢ 0.9000 −0.90 1 ⎥
⎥
˜ ⎢
B̃ = ⎢ 0.2700 ⎥
⎥
⎢ ⎥ ⎢ ⎥
(5.105)
⎢ 0.8075 −0.85 1 ⎥ ⎢ 0.2826 ⎥
⎣ ⎦ ⎣ ⎦

Thus, the first column of the matrix X is given by, see (5.48):

⎡ −1.6000 ⎤
⎢ ⎥
⎢ ⎥
X =⎢
⎢
⎥
⎥
⎢ ⎥
−2.7075 (5.106)
⎢ −0.7268 ⎥
⎣ ⎦

Then, the matrix A∗e is calculated as, cf. (5.51b):

⎡ −0.3912 ⎤T
⎢ ⎥
⎢ ⎥
⎢ 0.2150 ⎥
⎢ ⎥
∗ ⎢ ⎥
⎢
Ae = ⎢ −3.5079 ⎥⎥
⎢ ⎥
(5.107)
⎢ 12.7937 ⎥
⎢ ⎥
⎢ ⎥
⎢ −26.1909 ⎥
⎣ ⎦

Furthermore, matrices A′ and C ′ are obtained using (5.51c) and (5.51d):

⎡ 0.0019 ⎤
⎢ ⎥
⎢ ⎥
−0.0021 1.0106 −0.0192 −0.1078
⎢ −0.0010 ⎥
⎢ ⎥
⎢ ⎥
0.0012 −0.0058 1.0105 0.0593
′ ⎢
A =⎢ 0.0166 ⎥
⎥
⎢ ⎥
−0.0188 0.0954 −0.1720 0.0331
⎢ 0.9394 ⎥
⎢ 0.0686 −0.3480 0.6275 3.5263 ⎥
⎢ ⎥
(5.108)
⎢ −2.8760 ⎥
⎣ −0.2154 0.0779 −3.4220 −10.8064 ⎦
⎡ ⎤
⎢
′ ⎢ −0.0804 ⎥
C =⎢
0.0910 −0.4617 0.8324 −0.1968 ⎥
⎥
⎢ 0.0165 ⎥
⎣ ⎦
−0.0187 0.0947 −0.1708 0.0404

Similarly, as in the previous example:

⎡ ∗ ⎤
⎢ Ae ⎥
rank(CE) = rank(⎢
⎢
⎥) = 1
⎥
⎢ CE ⎥
(5.109)
⎣ ⎦

and (C ′ , A′ ) is a detectable pair, hence the solution to the robust fault detection prob-
lem exists. The remaining eigenvalues of (A − KC) are selected to be 0.45 and 0.4 and,

129
 5. Robust fault detection via eigenstructure assignment

consequently, the gain matrix K is given by, cf. (5.51e):

⎡ 2.6891 −0.4392 ⎤
⎢ ⎥
⎢ ⎥
⎢ 3.1731 −0.7126 ⎥
⎢ ⎥
⎢ ⎥
⎢
K = ⎢ 0.1983 0.9669 ⎥
⎥
⎢ ⎥
(5.110)
⎢ −0.7233 −3.5263 ⎥
⎢ ⎥
⎢ ⎥
⎢ 1.4808 ⎥
⎣ 7.2189 ⎦

The orthogonal projections of F on the subspace Im{Ē} and the orthogonal com-
pletion of Im{Ē} are given by:

⎡ 0.9687 0.1092 ⎤ ⎡ 0.0313 1.8908 ⎤

⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ −0.9910 −0.1749 ⎥ ⎢ −0.0090 3.5749 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢
F (Ē ) = ⎢ 1.0133 0.2406 ⎥ ⎢
F (Ē) = ⎢ −0.0133 1.7594 ⎥
⎥ ⎥
–
⎢ ⎥ ⎢ ⎥
(5.111)
⎢ 1.0324 0.0660 ⎥ ⎢ −0.0324 −0.0660 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.2718 ⎥ ⎢ 0.0282 −0.0125 ⎥
⎣ 0.0125 ⎦ ⎣ ⎦

One can calculate the generalised angles between, respectively, F1 and F2 and their
projections on the subspace Im{Ē}, cf. (5.102), which are, respectively, 88.4264o and
4.2009o . Note that the F1 is ‘almost orthogonal’ to the subspace Im{Ē}, whilst F2
‘almost lies’ in the subspace Im{Ē}, i.e. the angle between F2 and Im{Ē} is low. Since
the norm of F1 (Ē – ), i.e. 2.02, is significantly larger than the norm of F2 (Ē – ), i.e. 0.32,
it is expected that the steady state gain of the residual response to the fault µ1 (t) will
be larger than the one of µ2 (t). The responses of the residual to both faults in the
z-domain are given by:

r(z) 0.39613(z − 0.0221)

µ1 (z) (z − 0.45)(z − 0.4)
=
(5.112)
r(z) 0.094264(z − 0.08207)
µ2 (z) (z − 0.45)(z − 0.4)
=

where r(z), µ1 (z), and µ2 (z) are, respectively, z-domain representations of r(t), µ1 (t),
and µ2 (t). The responses of the residual to both faults are presented in Fig. 5.2. It can
be noted that the residual is insensitive to disturbances. The steady state gain of the
residual response to fault µ1 (t) is larger than the one of the residual response to fault
µ2 (t).

Example 5.3. q = 2, two invariant zeros

In this example q = 2 and there are two invariant zeros of the triple (A, E, C). The

130
 5. Robust fault detection via eigenstructure assignment

Disturbance
0.2

0.1
d(t)

−0.1

−0.2
0 10 20 30 40 50 60 70 80 90 100

Fault detection
µ1 (t)
1.5

µ2 (t)
µ1,2 (t), r(t)

1
r(t)
0.5

−0.5
0 10 20 30 40 50 60 70 80 90 100
Time [samples]

Figure 5.2: Robust fault detection, q = 1, the triple (A, E, C) has two invariant zeros
at 1.2 and 1.3. Upper subfigure presents disturbances, whilst lower sub-
figure demonstrates robust fault detection process, i.e. residual, r(t), is
insensitive to disturbances.

system is described by equation (5.1), whose matrices are:

⎡ 0 ⎤ ⎡ −0.8 0 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
0 1 0 0
⎢ ⎥ ⎢ 1.4 1 ⎥
⎢ 0 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
0 0 1 0
A=⎢⎢
⎥
0 ⎥ E=⎢⎢ 1.2 0 ⎥
⎥
⎢ ⎥ ⎢ ⎥
0 0 0 1
⎢ ⎥ ⎢ 0 1 ⎥
⎢ 0 0 0 0 1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ −0.0751 −0.6345 −2.1375 −3.5875 −3 ⎥ ⎢ 3.4816 1 ⎥
(5.113)
⎣ ⎦ ⎣ ⎦
⎡ 0.8675 0 −0.0837 ⎤
⎢ ⎥ ⎡ ⎤T
⎢ ⎥ ⎢ 1 1 0 0 0 ⎥
0.0948 −0.4811
C =⎢
⎢
⎥
⎥ F =⎢
⎢
⎥
⎥
⎢ ⎥ ⎢ 0 0 1 1 0 ⎥
0 0 0 1 0
⎢ 0.3929 ⎥ ⎣ ⎦
⎣ 0.8375 −0.3069 −0.2238 0 ⎦

Note that CE1 = 0 and δ1 = 1, whilst δ2 = 0. The invariant zero of (A, E1 , C) is 1.2 and
e1 is given by:
e1 = AE = [ 1.4 1.2 0 3.4816 −13.8381 ]
T
(5.114)

whilst e2 = E2 . Note that the triple (A, e1 , C) has two invariant zeros z1 = 0 and z2 = 1.2
= − [ 1 2 1 0 0 ]. Eigenvalues
(1) (1) (1)
and the corresponding v0 = e1 , v 1 = E1 and v2
of X (1) are selected to be 0.4, 0.5, and 0.6, whilst the eigenvalue corresponding to E2
is 0.7. Then matrices Ø(1) and B̃
˜ (1) are calculated as:

⎡ −0.32 −0.8 1 ⎤ ⎡ −0.128 ⎤

⎢ ⎥ ⎢ ⎥
⎢ ⎥ ˜=⎢ ⎥
Ø = ⎢
⎢
⎥
⎥
⎢ −0.175 ⎥
⎢ ⎥
⎢ ⎥ ⎢ ⎥
−0.35 −0.7 1 B̃ (5.115)
⎢ −0.36 −0.6 1 ⎥ ⎢ −0.216 ⎥
⎣ ⎦ ⎣ ⎦

131
 5. Robust fault detection via eigenstructure assignment

Thus, the first column of the matrix X is given by:

⎡ 0.300 ⎤
⎢ ⎥
⎢ ⎥
X (1) = ⎢
⎢
⎥
⎥
⎢ ⎥
−0.380 (5.116)
⎢ 0.336 ⎥
⎣ ⎦

Subsequently, the matrix A∗e is obtained using (5.62b) and (5.62c):

⎡ ⎤
⎢ ⎥
⎢ ⎥
0.1400 1
⎢ −0.5000 −0.7 ⎥
⎢ ⎥
∗ ⎢ ⎢
⎥
⎥
Ae = ⎢ ⎥
⎢ ⎥
3.6016 1 (5.117)
⎢ −14.8826 0.3 ⎥
⎢ ⎥
⎢ ⎥
⎢ 33.6320 −7.922 ⎥
⎣ ⎦

⎡ ∗ ⎤
Note that:
⎢ Ae ⎥
rank(CE) = rank(⎢
⎢
⎥) = 2
⎥
⎢ CE ⎥
(5.118)
⎣ ⎦
and (C ′ , A′ ) is a detectable pair, hence the solution to the robust fault detection prob-
lem exists. The remaining eigenvalue of (A − KC) is chosen to be 0.3. As a result the
gain matrix K is obtained as:

⎡ −0.7707 0.2586⎤
⎢ ⎥
⎢ ⎥
0.5424
⎢ 0.6203 −0.3448 −0.0556⎥
⎢ ⎥
⎢ ⎥
⎢
K = ⎢ −0.6264 0.6775 −0.3647⎥⎥
⎢ ⎥
(5.119)
⎢ −1.3397 −0.6720 ⎥
⎢ 2.5006 ⎥
⎢ ⎥
⎢ 6.5303 −3.4705 −8.8598⎥
⎣ ⎦

Note that the unobservable subspace of (QC, A − KC) is 4-dimensional, therefore

its observable subspace is only one-dimensional. This means that F1 (Ē – ) and F2 (Ē – )
are colinear and are given by:

⎡ 0.4397 ⎤ ⎡ 0.4397 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ −0.5181 ⎥ ⎢ −0.5181 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
– ⎢ ⎥
F1 (Ē ) = −0.4143 ⎢ 0.5965 ⎥ F2 (Ē ) = 1.0108 ⎢ 0.5965 ⎥
– ⎢
⎥
⎢ ⎥ ⎢ ⎥
(5.120)
⎢ 0.4143 ⎥ ⎢ 0.4143 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.1038 ⎥ ⎢ 0.1038 ⎥
⎣ ⎦ ⎣ ⎦

As a result the transfer functions between, respectively, µ1 (t) and µ2 (t) and the residual

132
 5. Robust fault detection via eigenstructure assignment

differ only by the steady state gain:

r(z) −0.29
µ1 (z) (z − 0.3)
=
(5.121)
r(z) 0.70
µ2 (z) (z − 0.3)
=

The results of the simulation are presented in Fig. 5.3. It can be seen that transient

Disturbance
d1 (t)
0.5

d2 (t)
d(t)

−0.5
0 10 20 30 40 50 60 70 80 90 100

Fault detection
1
µ1 (t)
µ2 (t)
µ1,2 (t), r(t)

0.5 r(t)
0

−0.5
0 10 20 30 40 50 60 70 80 90 100
Time [samples]

Figure 5.3: Robust fault detection, q = 2, the triple (A, E, C) has two invariant ze-
ros. Upper subfigure presents disturbances, whilst lower subfigure demon-
strates robust fault detection process. Trajectory of disturbances remains
within a 4-dimensional subspace of 5-dimensional state space of fault de-
tection filter, thus only one-dimensional subspace is left for fault detec-
tion. As a result the responses of residual to different faults differ only
by steady state gain, cf. Remarks 5.1 and 5.2.

behaviour of the residual responses to both faults is the same, but their steady state
gains differ.

5.3 Design of robust parity equations using right eigen-

structure assignment
In this section a novel design of robust PE is proposed. An illustration of the proposed
scheme is presented in Fig. 5.4. The method utilises a finite-time convergent observer
in order to obtain the state estimate. Then, by multiplying the estimated state vector,
x̂(t), by the system output matrix C and adding Du(t) an output estimate, ŷ(t), is
constructed and compared with the measured output. Note that the difference between
the robust residual generator described in the previous section and the scheme proposed
in this section is that the algorithm proposed here utilises a finite-time convergent

133
 5. Robust fault detection via eigenstructure assignment

state observer. This means that the state estimate converges within a finite time (as
opposed to the asymptotic observer in Subsection 2.6.1 and in the previous section).
Consequently, due to a finite impulse response of the state observer, the proposed
scheme is equivalent to PE. The material presented in this section is extension to
(Sumislawska, Larkowski & Burnham 2011b), where a design of robust PE for systems,
which do not contain any invariant zeros, has been proposed.

d(t) f (t)
u(t) y(t)

x̂(t) + ŷ(t) + r(t)

C + _ Q

Figure 5.4: Schematic illustration of the proposed residual generator

The finite-time convergent observer is presented in Subsection 5.3.1. Then in Sub-

section 5.3.2 the robust fault detection filter, which has been proposed in Section 5.2,
is combined with the finite-time convergent observer. In Subsection 5.3.3 it is demon-
strated that the proposed method is equivalent to PE. Finally, the design scheme is
explained using a numerical example in Subsection 5.3.4.

5.3.1 Finite-time convergent observer

A finite time convergent observer was originally developed by Engel & Kreisselmeier
(2002) for continuous-time systems. The scheme proposed here utilises an equivalent
observer in a discrete-time domain, which converges in a predefined time of τ samples.
For sake of completeness, the discrete-time form of the observer described by Engel &
Kreisselmeier (2002) is given in this subsection.
The system (5.1) in a fault-free, disturbance-free condition is described by:

x(t + 1) = Ax(t) + Bu(t)

(5.122)
y(t) = Cx(t) + Du(t)

Consider two Luenberger-type full-order state observers in the form of:

zi (t + 1) = Aci zi (t) + Ki y(t) + (B − Ki D)u(t) for i = 1, 2 (5.123)

where:
Aci = A − Ki C (5.124)

134
 5. Robust fault detection via eigenstructure assignment

The ith state estimation error ξi (t) = zi (t) − x(t) is then governed by:

ξi (t + 1) = Aci ξi (t) (5.125)

Therefore, the state estimation error as a function of time and initial conditions is
defined by:
ξi (t) = Atci ξi (0) (5.126)

The estimation error delayed by τ samples is then:

ξi (t − τ ) = At−τ
ci ξi (0) (5.127)

Subsequently, using (5.126) and (5.127), it holds that:

ξi (t) − Aτci ξi (t − τ ) = Atci ξi (0) − Atci ξi (0) = 0 (5.128)

thus ξi (t) can be eliminated from the state estimation zi (t):

zi (t) − Aτci zi (t − τ ) = x(t) − Aτci x(t − τ ) (5.129)

In order to obtain the correct state estimate the term Aτci x(t−τ ) needs to be eliminated
from the expression (5.129). Note that there are two unknowns in (5.129), namely,
x(t) and x(t − τ ), and two equations, i.e. two Luenberger-type observers, therefore an
explicit solution to the set of equations can be found. By combining the two observer,
the following is obtained, cf. (Engel & Kreisselmeier 2002):

z(t + 1) = Ac z(t) + Ky(t) + Gu(t) (5.130a)

x̂(t) = J (z(t) − Ac τ z(t − τ )) (5.130b)

where:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ Ac1 0 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
Ac = ⎢
⎢
⎥ K = ⎢ K1 ⎥ G = ⎢ B − K1 D
⎥ ⎢ ⎥ ⎢
⎥ z = ⎢ z1 ⎥
⎥ ⎢ ⎥
⎢ 0 Ac2 ⎥ ⎢ K2 ⎥ ⎢ B − K2 D ⎥ ⎢ z2 ⎥
(5.131)
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦

The gain matrix J should be chosen in such a way, the the terms Aτci x(t − τ ), i = 1, 2
are eliminated from (5.130b), i.e.

⎡ ⎤
⎢ x(t) − Aτc1 x(t − τ ) ⎥
(z(t) − Aτc z(t − τ )) =J⎢
⎢
⎥ = x(t)
⎥
⎢ x(t) − Aτc2 x(t − τ ) ⎥
J (5.132)
⎣ ⎦

Therefore, J should fulfil the following condition:

⎡ ⎤
⎢ I Aτc1 ⎥
J⎢
⎢
⎥=[ I 0 ]
⎥
⎢ I Aτc2 ⎥
(5.133)
⎣ ⎦

135
 5. Robust fault detection via eigenstructure assignment

Consequently, the gain matrix J is calculated as, cf. (Engel & Kreisselmeier 2002, Raff,
Menold, Ebenbauer & Allgöwer 2005):

⎡ ⎤−1
⎢ I Aτc1 ⎥
J = [ I 0 ]⎢
⎢
⎥
⎥
⎢ I Aτc2 ⎥
(5.134)
⎣ ⎦

Note that the solution of the expression (5.134) exists if the observer transition matrices
Ac1 and Ac2 have distinct eigenvalues. Therefore, if the system (A, E, C) contains an
invariant zero, Algorithm 5.3 should be used in order to ensure distinct eigenvalues of
Ac1 and Ac2 .

5.3.2 Proposed scheme

The necessary and sufficient conditions for a disturbance decoupling in the proposed
scheme are:

1. QCe = 0

2. For each column of the matrix E it holds that Im{e¯i } is an invariant subspace of
Ac1 and Ac2

3. Ac1 and Ac2 have distinct eigenvalues and no common eigenvalues

Conditions 1 and 2 ensure disturbance decoupling, whilst the condition 3 is essential

for an existence of a finite time convergent observer (Engel & Kreisselmeier 2002).
If the disturbance and fault vectors are present in the system (5.1), the state esti-
mation term of the ith observer, ξi (t) = zi (t) − x(t), see (5.127), is driven by, cf. (5.11)
and (5.12):

ξi (t + 1) = Aci ξi (t) + F µ(t) + Ed(t) = Aci ξi (t) + F µ(t) + ed∗ (t) (5.135)

Therefore, ei (t) can be defined by:

t−1 t−1
ξi (t) = Atci ξi (0) + ∑ Ajci F µ(t − j) + ∑ Ajci ed∗ (t − j) (5.136)
j=0 j=0

136
 5. Robust fault detection via eigenstructure assignment

Hence, the term zi (t) − Aτci z(t − τ ) can be expanded as:

t−1 t−1
zi (t) − Aτci z(t − τ ) = x(t) + Atci ξi (0) + ∑ Ajci F µ(t − j) + ∑ Ajci ed∗ (t − j)−
j=0 j=0
t−τ −1 t−τ −1
Aτci x(t − τ ) − Aτci At−τ
ci ξi (0) − Aci ∑ Aci F µ(t − τ − j) − Aci ∑ Aci ed (t − τ − j)
τ j τ j ∗
j=0 j=0
t−1 t−1 t−1
(5.137)
= x(t) − Aτci x(t − τ ) + ∑ Aci F µ(t − j) + ∑ Aci ed (t − j) − ∑ Aci F µ(t − j)−
j j ∗ j
j=0 j=0 j=τ
t−1 τ −1 τ −1
∑ Aci ed (t − j) = x(t) − Aci x(t − τ ) + ∑ Aci F µ(t − j) + ∑ Aci ed (t − j)
j ∗ τ j j ∗
j=τ j=0 j=0

Therefore, one can obtain the state estimate x̂(t), see (5.133):

⎡ τ −1 j ⎤ ⎡ τ −1 j ∗ ⎤
⎢ ∑j=0 Ac1 F µ(t − j) ⎥ ⎢ ⎥
x̂(t) = x(t) + J ⎢ ⎥ + J ⎢ ∑j=0 Ac1 ed (t − j) ⎥
⎢ τ −1 j ⎥ ⎢ τ −1 j ∗ ⎥
⎢ ∑j=0 Ac2 F µ(t − j) ⎥ ⎢ ∑j=0 Ac2 ed (t − j) ⎥
(5.138)
⎣ ⎦ ⎣ ⎦

Hence, the residual in the fault-free case is given by:

⎡ τ −1 j ∗ ⎤
⎢ ∑j=0 Ac1 ed (t − j) ⎥
r(t) = QCJ ⎢
⎢ τ −1 j ∗
⎥
⎥
⎢ ∑j=0 Ac2 ed (t − j) ⎥
(5.139)
⎣ ⎦

The matrix J can be expressed as:

J = [ J1 J2 ] (5.140)

where J1 , J2 ∈ Rn×n . Incorporating (5.140) into (5.134), the following relationships are
obtained:

J1 = −Aτc2 [Aτc1 − Aτc2 ]

−1
(5.141)
J2 = I − J1

Consequently, the residual in the fault-free case is:

⎡ τ −1 j ∗ ⎤
⎢ ∑j=0 Ac1 ed (t − j) ⎥
r(t) = QC [ J1 J2 ] ⎢
⎢ τ −1 j ∗
⎥
⎥
⎢ ∑j=0 Ac2 ed (t − j) ⎥
⎣ ⎦ (5.142)
τ −1 τ −1
= QCJ1 ∑ Ajc1 ed∗ (t − j) + QCJ2 ∑ Ajc2 ed∗ (t − j)
j=0 j=0

Denote a subspace spanned by columns of an arbitrary matrix V as V. Assume that

there exist two matrices P and R, such that V is P -invariant and R-invariant. From

137
 5. Robust fault detection via eigenstructure assignment

the definition of invariance it holds that (Halmos 1958):

PV ⊆ V
(5.143)
RV ⊆ V

Therefore:
P V + RV = (P + R)V ⊆ V (5.144)

This means that V is (P + R)-invariant. Furthermore, V is a P i -invariant subspace,

where i ∈ Z. This leads to the conclusion that Im{ē} is J1 - and J2 -invariant, cf. (5.141).
Due to the fact that Im{e} ⊆ Im{ē}, it holds that:

Ajci Im{e} ⊆ Im{ē} for i = 1, 2; j = 0, ⋯, τ − 1 (5.145)

because Im{ē} is Ajci -invariant and Im{e} ⊆ Im{ē}. Hence:

Ji Ajci Im{e} ⊆ Im{ē} for i = 1, 2; j = 0, ⋯, τ − 1 (5.146)

because Im{ē} is Ji -invariant and Im{e} ⊆ Im{ē}. Consequently:

QCJi Ajci Im{e} ⊆ QCIm{ē} = Im{QC ē} = 0 for i = 1, 2; j = 0, ⋯, τ − 1 (5.147)

Thus in the fault-free case the residual, cf. (5.139), is equal to zero.

5.3.3 Design of robust PE

A state estimation using the finite time convergent observer described in Subsection
5.3.1 can be expressed as a function of the last τ past values of the system input and
output:

2 τ −1 2 τ −1
x̂(t) = ∑ Ji ∑ Ajci (B − Ki D)u(t − j − 1) + ∑ Ji ∑ Ajci Ki y(t − j − 1) (5.148)
i=1 j=0 i=1 j=0

(The derivation of the above equation is analogous to (5.137).) Therefore, the residual
generator can be described by the following parity relation:

r(t) = Qy(t) − Q(C x̂(t) + Du(t)) = Wy Y (t) − Wu U (t) (5.149)

where:

Y (t) = [y T (t − τ ) y T (t − τ + 1) ⋯ y T (t)]
T

(5.150)
U (t) = [uT (t − τ ) uT (t − τ + 1) ⋯ uT (t)]
T

138
 5. Robust fault detection via eigenstructure assignment

and:

Wu = − [ QC(J1 Aτc1−1 (B − K1 D) + J2 Aτc2−1 (B − K2 D)) ⋯

QC(J1 A2c1 (B − K1 D) + J2 A2c2 (B − K2 D))

QC(J1 Ac1 (B − K1 D) + J2 Ac2 (B − K2 D))B QCB −QD ] (5.151)
Wy = [ QC(J1 Aτc1−1 K1 + J2 Aτc2−1 K2 ) ⋯ QC(J1 A2c1 K1 + J2 A2c2 K2 )

QC(J1 A1c1 K1 + J2 A1c2 K2 ) QC(J1 K1 + J2 K2 ) −Q ]

The algorithm for obtaining vectors Wu and Wy is given below:

Algorithm 5.4 (Robust PE via right eigenstructure assignment).

1. Obtain the disturbance direction matrix e = [ e1 e2 ⋯ eq ], where

ei = Aδi Ei (5.152a)

and δi is the smallest number for which CAδi Ei ≠ 0

2. Obtain Q such that QCe = 0

3. Select eigenvalues for Ac1 = A − K1 C

4. For each column of E obtain invariant zeros of the triple (A, ei , C), denoted
(i) (i)
as zj , and corresponding vectors vj , for j = 1, ⋯, qi , such that:

⎡ (i) ⎤ ⎡ (i) ⎤
⎢ zj I − A vj−1 ⎥ ⎢ vj ⎥
⎢ ⎥⎢ ⎥=0
(i)

⎢ ⎥⎢ ⎥
⎢ ⎥ ⎢ −1 ⎥
(5.152b)
⎣ C 0 ⎦⎣ ⎦
(i)
where v0 denotes ei .
(i) (i)
5. Calculate coefficients x11 , ⋯, xqi +1,1 using (5.59–5.61)

6. Obtain matrix A∗e as:

A∗e = [ A∗(1)
e
∗(2)
Ae
∗(q)
⋯ Ae ] (5.152c)

where

= (A − x11 I)ei − x21 v1 − x31 v2 − ⋯ − xqi +1,1 vk

(i) (i) (i) (i) (i) (i) (i)
A∗(i)
e (5.152d)

139
 5. Robust fault detection via eigenstructure assignment

7. Obtain:

A′ = A − A∗e (CE) C (5.152e)

C ′ = (I − CE(CE) ) C (5.152f)

8. Using any eigenstructure assignment method allocate remaining eigenvalues

of Ac1 = (A′ − K1′ C ′ )

9. Calculate K1
K1 = A∗e (CE) + K1′ [I − CE(CE) ] (5.152g)

10. Repeat steps 3 to 9 for Ac2 = A − K2 C

11. Choose τ and calculate J1 and J2 using

J1 = −Aτc2 [Aτc1 − Aτc2 ]

−1
(5.152h)
J2 = I − J1 (5.152i)

12. Obtain Wu and Wy using (5.151)

13. Calculate residual via

r(t) = Wy Y (t) − Wu U (t) (5.152j)

5.3.4 Numerical example

In this example Algorithm 5.4 is used in order to design robust PE. The influence of
the selection of eigenvalues of Ac1 and Ac2 and the convergence time τ on step and
impulse response of the residual to the fault is examined.

Example 5.4. Design of robust PE using right eigenstructure assignment

Consider the system (5.1), where:

⎡ 0 3 4 ⎤ ⎡ 1 0 ⎤ ⎡ 1 ⎤ ⎡ 1 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎡ ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 0 1 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢
A = ⎢ 1 2 3 ⎥ B = ⎢ 0 0 ⎥ E = ⎢ 2 ⎥ F = ⎢ −0.5 ⎥ C = ⎢ ⎥ (5.153)
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 0 0 1 ⎥
⎢ 0 2 5 ⎥ ⎢ 0 1 ⎥ ⎢ −1 ⎥ ⎢ 0.5 ⎥ ⎣ ⎦
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦

and the matrix D is null. The eigenvalues of the matrix Ac1 are chosen to be λAc1 =
{0.9, 0.925, 0.95}, whereas the eigenvalues of Ac2 are λAc2 = {0.965, 0.975, 0.995}. (Note
that eigenvalues of Ac1 and Ac2 are close to unity.) Gain matrices K1′ and K2′ have

140
 5. Robust fault detection via eigenstructure assignment

been selected to minimise Frobenius norms of K1 and K2 and are given by:

⎡ 3.6306 6.1612 ⎤ ⎡ 3.7124 6.3899 ⎤

⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢
K1 = ⎢ −0.3713 −0.9425 ⎥ K2 = ⎢ −0.4854 −1.0407 ⎥
⎥ ⎢
⎥
⎢ ⎥ ⎢ ⎥
(5.154)
⎢ 2.2481 ⎥ ⎢ 4.5504 ⎥
⎣ 4.5963 ⎦ ⎣ 2.2577 ⎦

Algorithm 5.4 is used to calculate Wu and Wy for parity space orders equal to τ = 5
and τ = 15 samples. The impulse and step responses of the residual r(t) to the fault
µ(t) for the two aforementioned parity space orders are compared in Fig. 5.5. It is
worth noting that the response of the residual to fault is strongly dependent on the
chosen parity space order. This is due to the fact that the eigenvalues of Ac1 and Ac2
are selected to be close to unity, therefore the two asymptotic state observers which are
combined to create the finite time convergent observers are relatively slow in comparison
to the chosen parity space orders.

Impulse response of residual to fault

τ =5
0.4

0.3 τ = 15
0.2

0.1

0
0 2 4 6 8 10 12 14 16 18 20
Step response of residual to fault
1

0.5

0
0 2 4 6 8 10 12 14 16 18 20
Time [samples]

Figure 5.5: Comparison of step and impulse responses of the residual to fault for
different cases of τ . Eigenvalues of (A − K1 C) and (A − K2 C) are, respec-
tively, λAc1 = {0.9, 0.925, 0.95} and λAc2 = {0.965, 0.975, 0.995}.

In the second simulation ‘fast’ eigenvalues are taken into consideration, i.e. λ∗Ac =
{0.33, 0.25, 0.3} and λ∗Ac = {0.15, 0.20, 0.35}, and compared with ‘slow’ eigenvalues
1

2
from the previous experiment for parity space τ = 15. The gain matrices K1∗ and K2∗
are:
⎡ 2.7472 3.8243 ⎤ ⎡ 2.7472 3.8243 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
∗ ⎢ ⎥ ∗ ⎢
K1 = ⎢ 0.7943 0.2487 ⎥ K2 = ⎢ 0.7943 0.2487 ⎥
⎥
⎢ ⎥ ⎢ ⎥
(5.155)
⎢ 2.3278 5.3257 ⎥ ⎢ 2.3278 5.3257 ⎥
⎣ ⎦ ⎣ ⎦

Step and impulse responses of the residual to the fault for two different sets of eigen-

141
 5. Robust fault detection via eigenstructure assignment

values (‘slow’, i.e. λAc1 and λAc2 , versus ‘fast’, i.e. λ∗Ac and λ∗Ac ) are compared in
1 2
Fig. 5.6. As expected, the response of the residual to the fault is faster when the
eigenvalues of Ac1 and Ac2 are closer to the origin. The experiment also revealed that,
in the case of ‘fast’ eigenvalues, the increase of the parity space order has a negligible
influence on the residual response to the fault.

Impulse response of residual to fault

0.4
fast
0.3 slow
0.2

0.1

0
0 2 4 6 8 10 12 14 16 18 20
Step response of residual to fault
1

0.5

0
0 2 4 6 8 10 12 14 16 18 20
Time [samples]

Figure 5.6: Comparison of step and impulse responses of the residual to fault
for different choices of eigenvalues of Ac1 and Ac2 . Slow eigenvalues:
λAc1 = {0.9, 0.925, 0.95} and λAc2 = {0.965, 0.975, 0.995}. Fast eigenval-
ues: λ∗Ac = {0.33, 0.25, 0.3} and λ∗Ac = {0.15, 0.20, 0.35}.
1 2

5.4 Design of robust parity equations using left eigen-

structure assignment
The basic idea of the fault detection filter presented in this Section is similar to that
described in Section 5.3. The state estimate is obtained using the finite time convergent
observer (see Subsection 5.3.1). Then, based on the state estimate, an output estimate
is calculated an compared with the measured output. The difference between the
measured and estimated output is then multiplied by the matrix Q, see Fig. 5.4. The
difference between the filter designed in this section and that presented in Section 5.3
is that the state observer gains K1 and K2 are obtained using the left eigenstructure
assignment, see Subsection 2.6.3.

5.4.1 Design of robust PE

Decoupling conditions of the proposed scheme are:

142
 5. Robust fault detection via eigenstructure assignment

1. QCE = 0

2. All rows of QC are left eigenvectors of Ac1 and Ac2

3. Ac1 and Ac2 have distinct eigenvalues and no common eigenvalues

Condition 2. is achieved by the left eigenstructure assignment algorithm presented in

Subsection 2.6.3. Analogously, to the case presented in Section 5.3, the fault detection
filter can be reformulated as a parity equation (5.149), whose matrices Wu and Wy are
given by (5.151). Consequently, the algorithm for the design of robust PE is summarised
as follows:

Algorithm 5.5 (Robust PE via left eigenstructure assignment).

1. Calculate the matrix Q, which fulfils the condition QCE = 0

2. Select the desired eigenvalues of the matrix Ac1 , λi , i = 1, 2, ⋯, n

3. Compute P (λi ) as
P (λi ) = (λi I − AT )−1 C T (5.156a)

4. Calculate li∗ and wi∗ , i = 1, ⋯, q using

li∗ = P (λi )wi∗ (5.156b)

wi∗ = [P (λi )T P (λi )]−1 P (λi )T li (5.156c)

and check the assignability condition li∗ = li .

5. Select arbitrary wi , i = q + 1, ⋯, n and obtain li , i = q + 1, ⋯, n using

li = −(λi I − AT )−1 C T wi (5.156d)

6. Obtain L and W as

L = [ l1∗ ⋯ lq∗ lq+1 ⋯ ln ] (5.156e)

W = [ w1∗ ⋯ wq∗ wq+1 ⋯ wn ] (5.156f)

7. Calculate K1 via
K = −(L−1 )T W T (5.156g)

8. Repeat steps (2-7) for the matrix Ac2 and obtain K2

143
 5. Robust fault detection via eigenstructure assignment

9. Choose τ and calculate J1 and J2 using

J1 = −Aτc2 [Aτc1 − Aτc2 ]

−1
(5.156h)
J2 = I − J1 (5.156i)

10. Construct Wu and Wy as given by (5.151)

11. Calculate residual via

r(t) = Wy Y (t) − Wu U (t) (5.156j)

Theorem 5.3. Algorithm 5.5 ensures disturbance decoupling.

Proof. Assume that two arbitrary matrices P and R have a common eigenvector x
corresponding to eigenvalues λp and λr , respectively:

P x = λp x Rx = λr x (5.157)

then x is also an eigenvector of their sum or difference:

(P ± R)x = (λp ± λr )x (5.158)

Furthermore, P can be expressed in a Jordan form as:

P = V ΛV −1 (5.159)

where diagonal elements of Λ are eigenvalues of P , whereas eigenvectors of P are

appropriate columns of V . Then it is known that:

P i = V Λi V −1 (5.160)

which means that P and its ith power have common eigenvectors. Consequently, if
all rows of QC are left eigenvectors of Ac1 and Ac2 , then all rows of QC are also left
eigenvectors of J1 and J2 .
The residual in the fault-free case can be described by, see (5.142):

2 τ
r(t) = ∑ ∑ Pl,k d(t − k) (5.161)
l=1 k=1

where:
Pl,k = QCJl Ak−1
cl E (5.162)

144
 5. Robust fault detection via eigenstructure assignment

Denote the ith left eigenvector, right eigenvector, and the corresponding eigenvalue of
the matrix Ak−1
cl as, respectively, (liAc )T , viAc , and λA
i . Analogously, use the notation
c

(liJ )T , viJ , and λJi for the left eigenvector, right eigenvector, and the corresponding
eigenvalue of the matrix Jl . The first rq left eigenvectors of Ak−1 k−1
c1 , Ac2 , J1 , and J2 are
rows of QC. Then Pl,k can be expressed as:

rq n
cl E + QC ∑ λi vi (li ) Acl E
Pl,k = QC ∑ λJi viJ (liJ )T Ak−1 J J J T k−1
(5.163)
i=1 i=rq +1

The second element of the above expression is equal to zero because QCviJ = 0, for
i = rq + 1, ⋯, n (because rows of QC are first left eigenvectors of Jl ). Subsequently:

rq rq

i λj vj (lj ) vi (li ) E
Pl,k = QC ∑ ∑ λA c J J J T Ac Ac T
i=1 j=1
n rq (5.164)
∑ ∑ λi c λj vj (lj ) vi c (li c ) E
A J J J T A A T
+ QC
i=rq +1 j=1

The first element of the above expression is zero because (liAc )T is equal to ith row of
QC and QCE = 0. The second element is equal to zero because (ljJ )T viAc = (ljJ )T viJ = 0,
for i = rq + 1, ⋯, n, j = 1, ⋯, rq . This shows, therefore, that the residual is null if there is
no fault present in the system.

5.4.2 Numerical example

In the following example, the design of robust PE using left eigenstructure assignment is
presented. The influence of the choice of τ as well as eigenvalues of Ac1 and Ac2 on step
and impulse responses of the residual to fault is examined. Furthermore, the proposed
scheme is compared with the DRFDF, i.e. a first order parity equation presented in
Subsection 2.6.4.

Example 5.5. Design of robust PE using left eigenstructure assignment

Consider system (5.1), whose matrices A, B, C, E, and F matrices are, cf. (Chen &
Patton 1999):

⎡ 0.25 0 0 ⎤ ⎡ 0 ⎤ ⎡ 1 ⎤ ⎡ 1.0 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎡ ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 1 1 0 ⎥
⎢
A=⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢
0 ⎥ B = ⎢ 0 ⎥ E = ⎢ 1 ⎥ F = ⎢ 0.1 ⎥ C = ⎢ ⎥ (5.165)
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 0 1 1 ⎥
0.5
⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎣ ⎦
⎣ 0 0.375 ⎦ ⎣ 1 ⎦ ⎣ 0 ⎦ ⎣ 1.0 ⎦

whilst the feedthrough matrix is null.

Eigenvalues of the matrix Ac1 are selected to be λAc1 = {0.975, 0.985, 0.995}, whilst
eigenvalues of Ac2 are λAc2 = {0.9, 0.925, 0.95}. The matrix Q, such that QCE = 0 is
given by:
Q = [ 1 −2 ] (5.166)

145
 5. Robust fault detection via eigenstructure assignment

Thus QC = l1T = [ 1 −1 −2 ]. The eigenvalue corresponding to l1T is λ1 = 0.975 and

hence P (λ1 ) is calculated via (5.156a) as:

⎡ 1.3793 0 ⎤
⎢ ⎥
⎢ ⎥
P (λ1 ) = ⎢
⎢
⎥
⎥
⎢ ⎥
2.1053 2.1053 (5.167)
⎢ 0 ⎥
⎣ 1.6667 ⎦

and corresponding w1∗ = [ 0.7250 −1.2 ] , cf. (5.156c). It holds that P (λ1 )w1∗ = l1 ,
T

hence a complete decoupling can be achieved. The remaining eigenvalues of Ac1 are
λ2 = 0.985 and λ3 = 0.995 and the corresponding matrices P (λ2 ) and P (λ3 ) are obtained
using (5.156a):

⎡ 1.3605 0 ⎤ ⎡ 1.3423 0 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
P (λ2 ) = ⎢
⎢
⎥
⎥ P (λ3 ) = ⎢
⎢
⎥
⎥
⎢ ⎥ ⎢ ⎥
2.0619 2.0619 2.0202 2.0202 (5.168)
⎢ 0 ⎥ ⎢ 0 ⎥
⎣ 1.6393 ⎦ ⎣ 1.6129 ⎦

Vectors w2 and w3 can be freely chosen and they are selected to be w2 = [ 1 0 ] and
T

w2 = [ 0 1 ] . Then, the corresponding l2 and l3 are computed using (5.156d):

T

⎡ 1.3605 ⎤ ⎡ 0 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
l2 = ⎢
⎢
⎥
⎥ l3 = ⎢
⎢ 2.0202
⎥
⎥
⎢ ⎥ ⎢ ⎥
2.0619 (5.169)
⎢ 0 ⎥ ⎢ 1.6129 ⎥
⎣ ⎦ ⎣ ⎦

Consequently, matrices L and W are obtained via, respectively, (5.156e) and (5.156e)

⎡ 1 1.3605 0 ⎤
as:
⎢ ⎥ ⎡ ⎤
⎢ ⎥ ⎢ 0.725 1 0 ⎥
L=⎢ ⎥
⎢ −1 2.0619 2.0202 ⎥ W =⎢
⎢
⎥
⎥
⎢ ⎥ ⎢ −1.200 0 1 ⎥
(5.170)
⎢ −2 0 1.6129 ⎥ ⎣ ⎦
⎣ ⎦
and the gain matrix K1 calculated using (5.156g) is:

⎡ 0.7203 −5.8212 ⎤
⎢ ⎥
⎢ ⎥
K1 = ⎢
⎢
⎥
⎥
⎢ ⎥
−0.9603 3.8412 (5.171)
⎢ 1.2028 −5.4312 ⎥
⎣ ⎦

Analogously, the gain matrix K2 is calculated. Vectors w2 and w3 are selected such
that the Frobenius norm of K2 is minimised and K2 is given by:

⎡ −1.6037 −0.5727 ⎤
⎢ ⎥
⎢ ⎥
K2 = ⎢
⎢
⎥
⎥
⎢ ⎥
0.7037 0.1227 (5.172)
⎢ −0.8287 −0.8727 ⎥
⎣ ⎦

Then, the matrices J1 and J2 and the PE coefficient vectors Wu and Wy are cal-

146
 5. Robust fault detection via eigenstructure assignment

Impulse response of residual to fault

0.4
T = 14
d
0.2 T =5
d

−0.2
2 4 6 8 10 12 14 16
Step response of residual to fault
1

0.5

2 4 6 8 10 12 14 16
Time [samples]

Figure 5.7: Comparison of responses of residual to fault for different cases of par-
ity space order. Eigenvalues of Ac1 and Ac2 are, respectively, λAc1 =
{0.9, 0.925, 0.95} and λAc2 = {0.965, 0.975, 0.995}.

culated using, respectively, (5.156h), (5.156i), and (5.151) for the parity space orders
equal to τ = 5 and τ = 14 samples. Impulse and step responses of the residual r(t) to
the fault µ(t) for the two aforementioned values of τ are compared in Fig. 5.7. Simi-
larly, as in Example 5.4 a strong influence of the parity space order τ on the residual
response to fault can be observed (due to the fact that the eigenvalues of Ac1 and Ac2
are close to unity). Also the impulse response of the residual to the fault decays almost
linearly. In the second simulation ‘fast’ eigenvalues are taken into consideration, i.e.
λ∗Ac = {0.2, 0.3, 0.4} and λ∗Ac = {0.8, 0.7, 0.6}, and compared with ‘slow’ eigenvalues
1 2
from the previous experiment for the parity space order τ = 8. The gain matrices K1∗
and K2∗ of the ‘fast’ filter are:

⎡ 0.0010 −0.0620 ⎤ ⎡ −0.5790 −0.1020 ⎤

⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢
= ⎢ 0.0390 0.0820 ⎥ = ⎢ 0.1390 −0.1180 ⎥
⎢
K1∗ ⎥ K2∗ ⎥
⎢ ⎥ ⎢ ⎥
(5.173)
⎢ −0.0440 0.1030 ⎥ ⎢ −0.0840 −0.4170 ⎥
⎣ ⎦ ⎣ ⎦

Step and impulse responses of the residual to the fault for two different sets of eigen-
values (‘slow’, i.e. λAc1 and λAc2 , versus ‘fast’, i.e. λ∗Ac and λ∗Ac ) are compared in
1 2
Fig. 5.8. Similarly to the experiment in Example 5.4, the response of the residual to
the fault is faster when the eigenvalues of Ac1 and Ac2 are closer to the origin. Also
in the case of ‘fast’ eigenvalues, the increase of the parity space order has negligible
influence on the residual response to the fault.
Choice of a low parity space order or selection of eigenvalues close to the origin leads
to a fast reaction of the residual to a fault. However, in practice there is always noise
present in the system. Therefore, it may be required to increase the parity space order

147
 5. Robust fault detection via eigenstructure assignment

Impulse response of residual to fault

1
‘fast’
‘slow’
0.5

1 2 3 4 5 6 7 8 9 10
Step response of residual to fault
1.5

0.5

−0.5
1 2 3 4 5 6 7 8 9 10
Time [samples]

Figure 5.8: Influence of eigenvalues of matrices Ac1 and Ac2 on step and impulse
response of residual to fault for τ = 8. Slow eigenvalues: λAc1 =
{0.975, 0.985, 0.995} and λAc2 = {0.9, 0.925, 0.95}. Fast eigenvalues:
λ∗Ac = {0.2, 0.3, 0.4} and λ∗Ac = {0.8, 0.7, 0.6}.
1 2

for the purpose of minimising the effects of noise on the residual generator. In this ex-
periment the scheme proposed here is compared with the DRFDF, see Subsection 2.6.4,
which is equivalent to the PE, whose Wu and Wy coefficients are:

Wu = [ 0 2 ]
(5.174)
Wy = [ 1 −2 −0.25 0.75 ]

Algorithm 5.5 has been designed using the ‘slow’ eigenvalues set, i.e. λAc1 and λAc2 for
two cases of the time delay, τ = 3 and τ = 8. The PE coefficient vectors for τ = 3 are:

Wu = [ 0.28 0.60 0.95 0 ]

(5.175)
Wy = [ −0.03 0.10 0.06 −0.06 0.18 −0.24 0.47 −0.95 ]

whilst those coefficients for τ = 8 are:

Wu = [ 0.04 0.8 0.13 0.18 0.24 0.31 0.39 0.47 0 ]

Wy = [ −0.005 0.014 0.009 −0.008 0.023 −0.031 0.04 −0.058 (5.176)

0.06 −0.089 0.081 −0.124 0.106 −0.163 0.133 ]

Fig. 5.9 shows the efficacy of the developed algorithm and the DRFDF, when the
system output is affected by an additive, white, Gaussian, zero-mean measurement noise
with variance equal to 0.01. The DRFDF yields a fast reaction of the residual to a fault,

148
 5. Robust fault detection via eigenstructure assignment

1.5

0.5 f(t)
DRFDF
0 T =3
d
T =8
d
−0.5
40 50 60 70 80 90 100
Time [samples]

Figure 5.9: Fault detection in the case when the output is subjected to measurement
noise

however it is relatively sensitive to noise. The filter obtained using Algorithm 5.5 allows
for an increase of the parity space order, hence a reduction of the residual sensitivity to
noise. Nevertheless, a high parity space order results in slow response of the residual
to the fault.

5.5 Concluding remarks

The drawback of the robust fault detection filter of Chen & Patton (1999) is its in-
applicability to systems with unstable invariant zeros. In such a case the design of a
stable filter was infeasible. In this chapter an extension to the aforementioned robust
fault detection filter has been presented which ensures stability of the scheme in the
case when the system has unstable invariant zeros. It has been demonstrated that the
eigenstructure of the developed scheme is equivalent that of the fault isolation filter
proposed in (Chen & Speyer 2006a). However, the algorithm presented in this chapter
is computationally simpler.
Furthermore, a novel design of robust PE of a user defined order has been presented.
In the proposed fault detection scheme the traditional asymptotically convergent ob-
server is replaced by a state observer, which converges within a finite, user predefined
time. By selecting the time (the number of samples) after which the state observer
converges, the order of the parity space can be arbitrarily chosen. This is an extension
to (Patton & Chen 1991b), where the left eigenstructure assignment method has been
used to design a first order PE. The method proposed in this chapter utilises both the
left and right eigenstructure assignment to design PE and is applicable to systems with
unstable invariant zeros.
Design freedom of the novel algorithms has been demonstrated on numerical exam-
ples. The residual response to faults can be shaped by a selection of the parity space
order as well as the eigenvalues of the component state observers, which form the finite
time convergent (open-loop) state observer. Advantages of the proposed scheme in a
noisy environment have also been shown. By selecting the order of the parity space the

149
 5. Robust fault detection via eigenstructure assignment

residual sensitivity to noise can be adjusted.

As further work an optimisation algorithm for minimisation of the influence of noise
on residual is considered. It may also be worth exploring the applicability of the scheme
when equation (5.52) is not fulfilled, i.e. the disturbance direction vectors ei combine
with each other to create new invariant zeros.

150
Chapter 6

Fault isolation via diagonal PE

Nomenclature

A .................. state transition matrix in state-space model

Ac1 , Ac2 . . . . . . . . . . . . filter state transition matrices
(i)
Aλ , A∗e , A∗e , Aw . . . auxiliary matrices
˜
Ã, Ã . . . . . . . . . . . . . . . auxiliary matrices
′
A ................. auxiliary matrix
B .................. input matrix of the input in state-space model
˜ ..................
B̃ auxiliary matrix
C .................. output matrix in state-space model
C′ . . . . . . . . . . . . . . . . . auxiliary matrix
D .................. feedforward matrix of known input in state-space model
f .................. matrix of directions of elements of µ(t)
fi . . . . . . . . . . . . . . . . . . direction of µi
f¯ . . . . . . . . . . . . . . . . . . matrix built from matrices fi
f¯i . . . . . . . . . . . . . . . . . . matrix whose image is sum of image of fi and images of invariant zero direc-
tions of (A, fi , C)
F .................. input matrix of fault signal in state-space model
Fi . . . . . . . . . . . . . . . . . ith column of F
g, gi . . . . . . . . . . . . . . . . auxiliary scalar
I ................... identity matrix
J, J1 , J2 . . . . . . . . . . . . gain matrices
K, K1 , K2 , K ′ . . . . . . gain matrices
lj , lj∗ . . . . . . . . . . . . . . . transposes of left eigenvectors of filter state transition matrix
m .................. number of system outputs
M (t) . . . . . . . . . . . . . . stacked vector of last τ + 1 values of µ(t)
n .................. number of states in a state-space model
p ................... number of system inputs
P (λi ) . . . . . . . . . . . . . . auxiliary function of λi
P, R . . . . . . . . . . . . . . . auxiliary matrices
q ................... number of disturbance signals
Q .................. gain matrix
r ................... number of fault signals
r(t) . . . . . . . . . . . . . . . . residual
ri (t) . . . . . . . . . . . . . . . ith element of r(t)

151
 6. Fault isolation via diagonal PE

T .................. similarity transformation matrix

u(t) . . . . . . . . . . . . . . . . measured system input
U (t) . . . . . . . . . . . . . . . stacked vector of last τ + 1 values of u(t)
(i)
v, vj , vj . . . . . . . . . . . auxiliary vectors
Ve . . . . . . . . . . . . . . . . . matrix whose columns are eigenvectors of filter state transition matrix
(i) ′(i)
wj , wj , wj . . . . . . . right eigenvectors of filter state transition matrix
wj∗ . . . . . . . . . . . . . . . . . auxiliary vector
W ................. auxiliary matrix
Wu , Wy , W̊u , W̊y , Wµ parity matrices

x(t) . . . . . . . . . . . . . . . . state vector instate space model

x̂(t) . . . . . . . . . . . . . . . . state estimate
xi,j . . . . . . . . . . . . . . . . auxiliary scalars
X .................. auxiliary matrix
y(t) . . . . . . . . . . . . . . . . measured system output
Y (t) . . . . . . . . . . . . . . . stacked vector of last τ + 1 values of y(t)
zi . . . . . . . . . . . . . . . . . . system zero
zi (t) . . . . . . . . . . . . . . . state estimate
αi . . . . . . . . . . . . . . . . . auxiliary parameter
βi , β̄i . . . . . . . . . . . . . . . auxiliary parameter vector
δi . . . . . . . . . . . . . . . . . . auxiliary term
(i)
λj , λj . . . . . . . . . . . . . eigenvalue of filter state transition matrix
Λe . . . . . . . . . . . . . . . . . diagonal matrix whose diagonal elements are eigenvalues of filter state transi-
tion matrix
µ(t) . . . . . . . . . . . . . . . . fault signal
µi (t) . . . . . . . . . . . . . . . ith element of fault signal
µ̂(t) . . . . . . . . . . . . . . . . estimate of fault signal
Θ(i) , Θ̄(i) . . . . . . . . . . . auxiliary matrices
Ω .................. set of all invariant zeros of (A, e, C) or auxiliary matrix
Ωi . . . . . . . . . . . . . . . . . set of all invariant zeros of (A, ei , C)
τ .................. convergence time of finite time-convergent state observer, order of parity space
Ξ(t) . . . . . . . . . . . . . . . auxiliary matrix
Ψ .................. auxiliary matrix

Preliminary reading: Sections 2.2, 2.6, 2.7, 5.2, and 5.3.

6.1 Introduction
Directional residuals have been used in various industrial applications, such as induc-
tion motor drives (Campos-Delgado 2011), a class of linear networked control systems
(Chabir, Sauter & Keller 2009), neuro-fuzzy diagnosis of AC motors (Alexandru 2003,
Alexandru & Popescu 2004), and engine fault detection and isolation (Dutka, Javahe-
rian & Grimble 2009). Also a structured residual set has been applied for fault diagnosis
of many industrial systems, such as: a two non-interacting tank system (Bhattacharjee
& Roy 2010), a heat exchanger (Fagarasan & St. Iliescu 2008), an aircraft (Fravolini,
Brunori, Campa, Napolitano & La Cava 2009), and a Tennessee Eastman process ex-
ample (Xie, Zhang & Wang 2006, Ye, Shi & Liang 2011).

152
 6. Fault isolation via diagonal PE

Furthermore, Patton & Chen (1997) proposed a scheme for condition monitoring
and fault diagnoisis of a seawater pumping system in operation at the Nuclear Electric
Heysham 2 power station. Simultaneous sensor and actuator fault diagnosis on a water
treatment system has been presented in (Fragkoulis, Roux & Dahhou 2009). Lia &
Jengb (2010) demonstrated a fault detection isolation and identification filter for a
nonisothermal continuous stirred tank reactor.
This chapter is an extension of the Algorithm 5.4 to a fault diagnosis scheme. In
Section 6.3 a fault isolation algorithm utilising a directional residual set is devised,
whilst in Section 6.4 the scheme is extended to fault identification.

6.2 Problem statement

It is assumed that a linear, dynamic, discrete-time, time-invariant system can be rep-
resented by the following equations:

x(t + 1) = Ax(t) + Bu(t) + F µ(t)

(6.1)
y(t) = Cx(t) + Du(t)

where x(t) ∈ Rn is the system state vector, u(t) ∈ Rp and y(t) ∈ Rm are, respectively, the
system input and output, and µ(t) ∈ Rr is a fault signal. Matrices A, B, C, D, and F
are constant and have appropriate dimensions. It is assumed that (C, A) is observable
and F is of full column rank. The aim of this chapter is to provide algorithms for fault
isolation an identification.

6.3 Design of fault isolation filter

A schematic illustration of the proposed fault isolation and identification filter is pre-
sented in Fig. 6.1. Analogously to the filter presented in Chapter 5, the method utilises
a finite-time convergent observer in order to obtain the state estimate. Then, using
the estimated state vector an output estimate is calculated and compared with the
measured output. The difference between the measured and the estimated output is a
directional residual (used for fault isolation). The residual is then evaluated in order
to identify the fault, see Section 6.4.
Design of the fault isolation filter is analogous to that proposed in Section 5.3, i.e.
the fault isolation filter is designed using Algorithm 5.4 by replacing the matrix E by
F and setting Q to an identity matrix. The necessary conditions for the fault isolation
filter to exist are:

1. (C,A) is an observable pair

2. rank([ Cf1 Cf2 ⋯ Cfr ]) = r

153
 6. Fault isolation via diagonal PE

d(t) µ(t)
u(t) y(t)

x̂(t)
+ ŷ(t) + r(t) µ̂(t)
C + _ residual
evaluation

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶

fault isolation fault identification
Figure 6.1: Schematic illustration of the proposed residual generator

where fi = Aδi Fi and δi is the smallest number for which CAδi Fi ≠ 0. Condition 1
allows all eigenvalues of (A − K1 C) and (A − K2 C) to be arbitrarily specified, whilst
condition 2 ensures that residual vectors yielded by different faults lie in different di-
rections.

Algorithm 6.1 (Fault isolation using directional PE).

1. For each column of F obtain δi , which is the smallest number for which
CAδi Fi ≠ 0 and compute zero directions fi = Aδi Fi

2. Select eigenvalues for Ac1 = (A − K1 C)

3. For fi , i = 1, ⋯, r obtain invariant zeros of the triple (A, fi , C), denoted as zj ,

(i)

(i)
and corresponding directions vj , for j = 1, ⋯, ri , where ri is the number of
invariant zeros of the triple (A, fi , C), such that

⎡ (i) ⎤ ⎡ (i) ⎤
⎢ zj I − A vj−1 ⎥ ⎢ vj ⎥
⎢ ⎥⎢ ⎥=0
(i)

⎢ ⎥⎢ ⎥
⎢ ⎥ ⎢ −1 ⎥
(6.2a)
⎣ C 0 ⎦⎣ ⎦
(i)
where v0 denotes fi .

4. Obtain matrix Ø(i) and vector B̃

˜ (i) whose elements are given by

r
Øj,k = (−1)k−1 ∏(zl − λj )
i
(i) (i) (i)
(6.2b)
l=k
r
˜ (i) = λ i (z (i) − λ(i) )
B̃ j j∏ l j (6.2c)
l=1

154
 6. Fault isolation via diagonal PE

(i) (i)
5. For i = 1, ⋯, r calculate coefficients x11 , ⋯, xri +1,1 which fulfil the condition

[ x(i) x12 ⋯ xri +1,1 ] = (Ã(i) )−1 B̃ (i)

(i) (i) T ˜ ˜
11 (6.2d)

6. Obtain matrix A∗e as:

A∗e = [ A∗(1)
e
∗(2)
Ae
∗(r)
⋯ Ae ] (6.2e)

where

= (A − x11 I)fi − x21 v1 − x31 v2 − ⋯ − xri +1,1 vr(i)

(i) (i) (i) (i) (i) (i)
A∗(i)
e i
(6.2f)

7. Obtain:

A′ = A − A∗e (CE) C (6.2g)

C = (I − CE(CE) ) C
′
(6.2h)

8. Using any eigenstructure assignment methods allocate remaining eigenvalues

of Ac1 = (A′ − K1′ C ′ )

9. Calculate K1 as

K1 = A∗e (CE) + K1′ (I − CE(CE) ) (6.2i)

10. Repeat steps 2 to 9 for the gain matrix K2 and obtain Ac1 = A − K2 C

11. Choose τ and calculate J1 and J2 using

J1 = −Aτc2 [Aτc1 − Aτc2 ]

−1
(6.2j)
J2 = I − J1 (6.2k)

12. Calculate Wu and Wy via

Wu = − [ C(J1 Aτc1−1 (B − K1 D) + J2 Aτc2−1 (B − K2 D)) ⋯

C(J1 A2c1 (B − K1 D) + J2 A2c2 (B − K2 D)) (6.2l)

C(J1 Ac1 (B − K1 D) + J2 Ac2 (B − K2 D))B CB −D ]
Wy = [ C(J1 Aτc1−1 K1 + J2 Aτc2−1 K2 ) ⋯ C(J1 A2c1 K1 + J2 A2c2 K2 )

C(J1 A1c1 K1 + J2 A1c2 K2 ) C(J1 K1 + J2 K2 ) −I ] (6.2m)

155
 6. Fault isolation via diagonal PE

13. Calculate residual via

r(t) = Wy Y (t) − Wu U (t) (6.2n)

Remark 6.1. The residual ri (t) obtained using Algorithm 6.1 lies in the direction Cfi .

Demonstration. The residual calculated using Algorithm (6.1) is, cf. (5.139):

⎡ τ −1 j ⎤
⎢ ∑j=0 Ac1 F µ(t − j) ⎥
r(t) = C [ J1 J2 ] ⎢
⎢ τ −1 j
⎥
⎥
⎢ ∑j=0 Ac2 F µ(t − j) ⎥
⎣ ⎦ (6.3)
τ −1 τ −1
= CJ1 ∑ Ajc1 F µ(t − j) + QCJ2 ∑ Ajc2 F µ(t − j)
j=0 j=0

Consider a matrix f¯i = [ fi v1(i) ⋯ vr(i) i

]. It has been shown in Subsection 5.3.2,
that Im{f¯i } is an invariant subspace of Ac1 , Ac2 , J1 , and J2 . Hence, it holds that, cf.
Subsection 5.3.2:

CJk Ajck Im{f¯i } ⊆ CIm{f¯i } for k = 1, 2; j = 0, ⋯, τ − 1, i = 1, ⋯, r (6.4)

(i)
Recall that Cvj = 0 for j = 1, ⋯, rl , i = 1, ⋯, r. Consequently:

Im{C f¯i } = Im{Cfi } for i = 1, ⋯, r (6.5)

Thus, a design of the fault isolation filter using Algorithm 6.1 ensures that the residual
driven by the fault signal µi (t) lies in the direction Cfi .

6.4 Fault identification

6.4.1 Change of coordinates
The presence of the ith fault yields residual direction Cfi . Thus, it is useful to represent
the residual vector r(t) as a linear combination of residual directions Cfi , i = 1, 2, ⋯, r,
i.e.
r(t) = γ1 Cf1 + γ2 Cf2 + ⋯ + γr Cfr (6.6)

where γ1 , γ2 , ⋯, γr are scalar coefficients. The parameter γi deviating from zero indicates
a presence of the fault µi (t) and its value depends on the magnitude of the fault.
Consequently, a change of coordinates is defined as:

T = [ T 1 T2 ] (6.7)

156
 6. Fault isolation via diagonal PE

where T1 = [ Cf1 Cf2 ⋯ Cfr ] and Im{T2 } is an orthogonal completion of Im{T1 }.

Applying the above similarity transformation to r(t) a variable r̊(t) is obtained:

r̊(t) = T −1 r(t) (6.8)

Remark 6.2. An occurrence of µi (t) causes r̊i (t) deviate from zero, whilst the re-
maining components of r̊(t) are zero.

Demonstration. The term r̊(t) is given by:

r̊(t) = [ Cf1 Cf2 ⋯ Cfr T 2 ] (γ1 Cf1 + γ2 Cf2 + ⋯ + γr Cfr )

−1
(6.9)

which can be reformulated as:

[ Cf1 Cf2 ⋯ Cfr T 2 ]r̊(t) = γ1 Cf1 + γ2 Cf2 + ⋯ + γr Cfr (6.10)

which it true if and only if:

r̊(t) = [ γ1 γ2 ⋯ γr 0 ⋯ 0 ]
T
(6.11)

Thus r̊i (t) deviates from zero only when the fault µi (t) occurs.

6.4.2 Steady state gain calculation

From (6.3) it follows that the directional residual as a function of a fault signal can be
represented by:
r(t) = Wµ M (t) (6.12)

where:

Wµ = [ C(L1 Aτc1−1 + L2 Aτc2−1 )F ⋯ C(L1 A2c1 + L2 A2c2 )F

(6.13)
C(L1 Ac1 + L2 Ac2 )F QCF 0 ]

and:
M (t) = [ µT (t − τ ) µT (t − τ + 1) ⋯ µT (t) ]
T
(6.14)

It has been shown that the fault direction driven by µi (t) is Cfi . Consequently, the
directional residual, r(t), can be formulated as:
τ r
r(t) = ∑ ∑ αj Cfi µi (t − j)
(i)
(6.15)
j=0 i=1

157
 6. Fault isolation via diagonal PE

(i)
where αj are scalar coefficients. Then, the term r̊(t) can be reformulated as:

r̊(t) = T −1 Wµ M (t) = ΩM (t) (6.16)

where Ω ∈ Rm×k(τ +1) is given by Ω = T −1 Wµ .

Remark 6.3. The term Ω is in the form of:

⎡ (1) ⎤
⎢ ατ ⎥
(1) (1)
⎢ ⎥
0 ⋯ 0 ατ −1 0 ⋯ 0 ⋯ α0 0 ⋯ 0
⎢ 0 ⎥
⎢ ⎥
(2) (2) (2)
⎢ ⎥
ατ ⋯ 0 0 ατ −1 ⋯ 0 ⋯ 0 α0 ⋯ 0
⎢ ⋮ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮
⎢
Ω=⎢ 0 ⎥(6.17)
⎥
(k) (k) (k)
⎢ ⎥
0 ⋯ ατ 0 0 ⋯ ατ −1 ⋯ 0 0 ⋯ α0
⎢ 0 ⎥
⎢ ⎥
⎢ ⎥
0 ⋯ 0 0 0 ⋯ 0 ⋯ 0 0 ⋯ 0
⎢ ⋮ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ ⋱ ⋮ ⋮ ⋱ ⋮
⎢ 0 ⎥
⎣ 0 ⋯ 0 0 0 ⋯ 0 ⋯ 0 0 ⋯ 0 ⎦

Demonstration. See Appendix C

Note that r̊i (t) is the residual vector r(t) represented by the basis T , and, consequently,
r̊i (t) can be reformulated as, cf. (6.15):

r̊i (t) = α0 µi (t) + α1 µi (t − 1) + ⋯ + ατ(i) µi (t − τ )

(i) (i)
(6.18)

Therefore, the steady state gain of the response of r̊i (t) to the fault µi (t) is equal to
(i)
∑τj=0 αj , which is the sum of elements of the ith row of Ω. Consider a diagonal matrix:

Ξ = diag [ ∑τi=1 αi(1) ∑τi=1 αi(2) ⋯ ∑τi=1 αi(k) 1 ⋯ 1 ] (6.19)

and the variable:

µ̂(t) = Ξ−1 r̊(t) (6.20)

The first r elements of the vector µ̂(t) are estimates of µ(t), whereas remaining m − r
elements of µ̂(t) should be equal zero and may be treated as control variables (if they
deviate from zero it indicates there is a fault or a disturbance that is not covered by
the model). Consequently, the PE for fault isolation and identification is given by:

µ̂(t) = W̊y Y (t) + W̊u U (t) (6.21)

where

W̊u = Ξ−1 T −1 Wu and W̊y = Ξ−1 T −1 Wy (6.22)

The algorithm for fault isolation and identification via diagonal PE is summarised
below.

158
 6. Fault isolation via diagonal PE

Algorithm 6.2 (Fault isolation and identification via diagonal PE).

1. Compute r(t) using Algorithm 6.1

2. Obtain T using (6.7)

3. Obtain Wµ using (6.13) and obtain Ω as:

Ω = T −1 Wµ (6.23a)

4. Compute Ξ using (6.19)

5. Calculate W̊u and W̊y as:

W̊u = Ξ−1 T −1 Wu (6.23b)

W̊y = Ξ−1 T −1 Wy (6.23c)

6. Compute reconstructed fault vector as:

µ̂(t) = W̊y Y (t) + W̊u U (t) (6.23d)

6.5 Consideration of measurement noise

In practice there is noise present in the system. Consequently, the residual is rarely
equal to zero, and a decision must me made, whereas the residual deviating from zero
indicates presence of a fault or is a result of noise. A fault presence is sensed if a residual
exceeds a certain threshold (Ding 2008). An appropriate choice of the threshold allows
minimisation of the number of false alarms as well as missed fault occurrences. In this
subsection a simple method for calculating thresholds is presented.
Consider the system (6.1) in the EIV framework:

x(t + 1) = Ax(t) + Bu0 (t) + F µ(t)

y0 (t) = Cx(t) + Du(t)
(6.24)
u(t) = u0 (t) + ũ(t)
y(t) = y0 (t) + ỹ(t)

where noise-free input and output, u0 (t) and y0 (t) are affected by white, Gaussian,
zero-mean, mutually uncorrelated noise sequences ũ(t) and ỹ(t), respectively. The
terms u(t) and y(t) are measured values of the input and output. The term µ̂(t) is

159
 6. Fault isolation via diagonal PE

calculated as:

µ̂(t) = W̊y Y (t) + W̊u U (t) = W̊y (Y0 (t) + Ỹ (t)) + W̊u (U0 (t) + Ũ (t)) (6.25)

which in a fault-free case is:

µ̂ = W̊y Ỹ (t) + W̊u Ũ (t) (6.26)

It should be noted that each column of µ̂(t) is affected by the measurement noise in
different level. Therefore, thresholds calculation should be based on the expected values
of the variance each element of µ̂(t) in the fault-free case. The covariance matrix of
µ̂(t), denoted as Σµ̂ , in fault-free case is given by:

Σµ̂ = W̊y Σỹ W̊yT + W̊u Σũ W̊uT (6.27)

where Σỹ = E{Ỹ (t)Ỹ T (t)} and Σũ = E{Ũ (t)Ũ T (t)}. In the EIV framework with no
process noise, cf. (6.24), Σỹ and Σũ are diagonal matrices. Based on the assumption
that the measurement noise sequences are white, Gaussian, zero-mean, and mutually
uncorrelated, the threshold which µ̂i (t) needs to violate for the fault to be noticed
should be calculated as an appropriate multiplicity of the standard deviation of µ̂i (t)
in a fault-free case, i.e. an appropriate multiplicity of the square root of the ith diagonal
element of Σµ̂ .

6.6 Numerical example

Example 6.1. Design of diagonal PE for fault isolation and identification
It is assumed that a linear discrete-time time-invariant system is described by (6.24)
where the system matrices are:

⎡ ⎤ ⎡ 0 ⎤ ⎡ 1 1 0 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 1 0 0 0
⎢ ⎥ ⎢ ⎥ ⎢ 0 0 ⎥
⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ 1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 0 1 0
A=⎢⎢ 0 ⎥ B=⎢
⎥
⎢ 0 ⎥ F = ⎢ 0 −1 0 ⎥
⎥ ⎢
⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
0 0 0 1
⎢ 1 ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥
⎢ 0 0 0 0 ⎥ ⎢ ⎥ ⎢ 2 1 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ −4.85 3.5 ⎥ ⎢ 1 ⎥ ⎢ 0 ⎥ (6.28)
⎣ 0.1512 −1.1274 3.325 ⎦ ⎣ ⎦ ⎣ 0 2 ⎦
⎡ 1 −1 0 0 0 ⎤ ⎡ ⎤
⎢ ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 1 1 0 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
C =⎢ ⎥ D=⎢ ⎥
0
⎢ ⎥ ⎢ 0 ⎥
⎢ 0 1 0 1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
0
⎢ ⎥ ⎢ 0 ⎥
⎣ 0 0 0 0 1 ⎦ ⎣ ⎦

The first output is affected by white, Gaussian, zero-mean noise with the variance equal
to σy21 = 0.01, whilst the remaining three are subjected to white, Gaussian, zero-mean
noise sequences with the variances of σy22 = σy23 = σy24 = 0.0001. Output measurement

160
 6. Fault isolation via diagonal PE

noise sequences are mutually uncorrelated.

Note that CF1 = 0 and:

f1 = AF1 = [ 1 0 0 0 −0.9762 ]
T
(6.29)

Therefore, δ1 = 1. Due to the fact that CF2 ≠ 0 and CF3 ≠ 0, f2 = F2 and f3 = F3 .

Consequently, residual directions are:

⎡ 1 ⎤ ⎡ 1 ⎤ ⎡ 0 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ ⎥ ⎢ 1 ⎥
⎢ ⎥ ⎢ 1 ⎥ ⎢ ⎥
Cf1 = ⎢ ⎥ Cf2 = ⎢ ⎥ Cf3 = ⎢ ⎥
⎢ −0.9762 ⎥ ⎢ ⎥ ⎢ 2 ⎥
(6.30)
⎢ ⎥ ⎢ −1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ −0.9762 ⎥ ⎢ 0 ⎥ ⎢ 2 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦

Note that rank([ Cf1 Cf2 Cf3 ]) = 3, hence the possibility for a complete fault
isolation exists.
Eigenvalues of (A − K1 C) are chosen to be 0.22, 0.77, 0.66, 0.55, 0.44. Note that
the triple (A, f1 , C) has an invariant zero at z1
(1)
= 0, whilst the state and input zero
= −1. Eigenvalues of (A − K1 C) corre-
(1) (1)
directions are, respectively, v1 = F1 and g1
sponding to a linear combination of columns of f¯1 = [ f1 F1 ] are 0.22 and 0.44. As
a result, the first column of the matrix X (1) is given by, cf. (6.2b), (6.2c), (6.2d):

⎡ ⎤
⎢ 0.6600 ⎥
⎢
=⎢ ⎥
⎥
(1)
⎢ −0.0968 ⎥
X (6.31)
⎣ ⎦

Hence, A∗(1) is calculated as, cf. (6.2f):

A∗(1) = (A − 0.66I)f1 + 0.0968v1 = [ −0.5632 0 0 −0.9762 −2.2157 ]

(1) T
(6.32)

whereas A∗(2) and A∗(3) are:

A∗(2) = (A − 0.55I)f2 A∗(3) = (A − 0.66I)f2 (6.33)

Subsequently, the matrix A∗ is obtained as, cf. (6.2e):

⎡ −0.5632 −0.5500 ⎤
⎢ ⎥
⎢ ⎥
0
⎢ 0 ⎥
⎢ ⎥
∗ ⎢ ⎥
−1 0
⎢
A =⎢ 0 ⎥
⎥
⎢ ⎥
2.5500 1 (6.34)
⎢ −0.9762 −1.1000 1.34 ⎥
⎢ ⎥
⎢ ⎥
⎢ −2.7157 −12.8738 0.83 ⎥
⎣ ⎦

The remaining eigenvalue of (A′ − K1′ C ′ ) is chosen to be 0.77. Subsequently, the gain

161
 6. Fault isolation via diagonal PE

matrix K1 is computed as:

⎡ −0.4196 −0.2943 −0.1639 0.3110 ⎤

⎢ ⎥
⎢ ⎥
⎢ 0.2175 −0.2472 0.9702 −0.8466 ⎥
⎢ ⎥
⎢ ⎥
⎢
K1 = ⎢ −0.3623 0.7988 ⎥
⎥
⎢ ⎥
1.7423 −1.1700 (6.35)
⎢ 0.0020 −0.6641 0.5642 ⎥
⎢ 0.4379 ⎥
⎢ ⎥
⎢ −1.0702 −2.3477 ⎥
⎣ 9.4559 −7.8671 ⎦

The eigenvalues of (A − K2 C) are selected to be 0.95, 0.85, 0.18, 0.33, 0.47 and, as
a consequence, K2 is calculated as:

⎡ −0.9584 −0.0699 −0.5583 0.5932 ⎤

⎢ ⎥
⎢ ⎥
⎢ 1.0705 −0.5388 1.5317 −1.2623 ⎥
⎢ ⎥
⎢ ⎥
K2 = ⎢
⎢
⎥
⎥
⎢ ⎥
−0.3388 1.6940 −1.1147 0.7677 (6.36)
⎢ 0.1891 −0.5675 ⎥
⎢ 0.5616 0.6321 ⎥
⎢ ⎥
⎢ 1.1636 −3.6841 10.3533 −7.6162 ⎥
⎣ ⎦

The order of the parity space is chosen as τ = 8 and matrices Wu and Wy are
calculated using (6.2l) and (6.2m), respectively. Then T is obtained using (6.7):

⎡ 1 0.3072 ⎤
⎢ ⎥
⎢ ⎥
1 0
⎢ 0 1 1 −0.6294 ⎥
⎢ ⎥
T =⎢ ⎥
⎢ −0.9762 −1 2 −0.3222 ⎥
(6.37)
⎢ ⎥
⎢ ⎥
⎢ −0.9762 0 2 0.6369 ⎥
⎣ ⎦

The matrix Ω is then computed as, cf. (6.23a):

⎡
⎢
⎢
0 0 0 2.3e − 2 0 0
⎢
⎢
Ω = T Wµ = ⎢
0 3.2e − 2 0 0 1.3e − 1 0
⎢
−1

⎢
⎢
0 0 4.3e − 3 0 0 3e − 2
⎢
⎣ 0 0 0 0 0 0
(6.38)
1 0 0 0 0 0 ⎤
⎥
⎥
1.8e − 1 0 0
0 1 0 0 0 0 ⎥
⎥
⎥
0 3.8e − 1 0
1.8e − 1 0 0 1 0 0 0 ⎥
⎥
⎥
0 0
0 0 0 0 0 0 0 0 0 ⎥
⎦

The first three diagonal elements of Ξ are equal to sum of appropriate rows of Ω. The

162
 6. Fault isolation via diagonal PE

last diagonal element of Ξ is unity, see (6.19):

⎡ 0 ⎤
⎢ ⎥
⎢ ⎥
1.2004 0 0
⎢ 0 ⎥
⎢ ⎥
Ξ=⎢ ⎥
0 1.5386 0
⎢ 1.2118 0 ⎥
(6.39)
⎢ ⎥
⎢ ⎥
0 0
⎢ 1 ⎥
⎣ 0 0 0 ⎦

Consequently, W̊u and W̊y are calculated using, respectively, (6.23b) and (6.23c) and
the fault estimate is obtained using (6.23d).

Measurement noise

Recall that output measurements are affected by white, zero-mean, Gaussian, and
mutually uncorrelated noise sequences with variances of σy21 = 0.01 and σy22 = σy23 =
σy24 = 0.0001, whilst τ = 8. Hence, the output measurement covariance matrix is
a 36th order ((τ + 1)m = 9 × 4 = 36) diagonal matrix, whose diagonal is the vector
[ σy21 σy22 σy23 σy24 ] repeated τ + 1 = 9 times. Consequently, Σµ̂ is calculated as,
cf. (6.27):
⎡ 0.0190 −0.0119 0.0048 ⎤
⎢ ⎥
⎢ ⎥
0.0129
⎢ −0.0119 0.0190 −0.0118 −0.0034 ⎥
⎢ ⎥
Σµ̂ = ⎢ ⎥
⎢ 0.0129 −0.0118 0.0045 ⎥
(6.40)
⎢ ⎥
⎢ ⎥
0.0118
⎢ 0.0048 −0.0034 ⎥
⎣ 0.0045 0.0038 ⎦
Thus, the variances of the consecutive elements of µ̂(t) in a fault-free case are, respec-
tively, 0.0190, 0.0190, 0.0118, and 0.0038. The threshold of each element of µ̂(t) is
selected as its standard deviation in a fault-free case multiplied by 3.1 (which results in
the confidence bound of 0.999, assuming a Gaussian distribution of the measurement
noise) and are given by, respectively, 0.4272, 0.3557, 0.3363, and 0.1915. Results of the
simulation are presented in Fig. 6.2. The filter identifies abrupt faults µ1 (t) and µ2 (t),
as well as an incipient fault µ3 (t). After the 100th sample an unmodelled fault occurs.
This is indicated by µ̂4 (t) violating the threshold.

6.7 Conclusions
The fault isolation and identification filter devised in this chapter utilises a diagonal
residual set. Therefore, multiple faults can be isolated and identified. Furthermore, if
the number of linearly independent outputs exceeds the number of modelled faults, a
new variable has been introduced, which indicates occurrence of an unmodelled fault.
This provides a signal to stop the process and investigate the source of a possibly
dangerous fault. Also a straightforward method to calculate residual thresholds, whose
violation indicates a fault, has been proposed for the EIV framework. A simulation
study has demonstrated promising results for both diagnosis of abrupt and incipient

163
 6. Fault isolation via diagonal PE

Residual response to faults

f1 (t)
6
4
µ̂1 (t)
2
0
−2
10 20 30 40 50 60 70 80 90 100 110 120

f2 (t)
8
6
µ̂2 (t)
4
2
0
10 20 30 40 50 60 70 80 90 100 110 120

0
−2
−4
f3 (t)
µ̂3 (t)
−6
−8
10 20 30 40 50 60 70 80 90 100 110 120

−2

µ̂4 (t)
−4
10 20 30 40 50 60 70 80 90 100 110 120
Time [samples]

Figure 6.2: Fault identification using diagonal PE. At the 100th time sample unmod-
elled fault occurs causing µ̂4 (t) deviate from zero.

faults.
Further work aims to extend the scheme to fault isolation and identification of
multidimensional faults.

164
Chapter 7

Potential applications

Nomenclature

Ap . . . . . . . . . . . . . piston area
C .............. compensation variable
d1 , d2 , d2 . . . . . . . . damping coefficients
Fc (t) . . . . . . . . . . . Coulomb friction
Ff ric (t) . . . . . . . . total friction force
Fh (t) . . . . . . . . . . . hydraulic force
Fr (t) . . . . . . . . . . . roll force
Fs (t) . . . . . . . . . . . Stribeck friction (stiction)
Fv (t) . . . . . . . . . . . viscous friction
G1 , G2 . . . . . . . . . . controller gains
Gv (z) . . . . . . . . . . transfer function
k1 . . . . . . . . . . . . . . spring constant
k2 . . . . . . . . . . . . . . spring constant
k3 . . . . . . . . . . . . . . steel strip spring constant
Kc . . . . . . . . . . . . . hydraulic oil compressibility coefficient
Kp . . . . . . . . . . . . . proportional gain of hydraulic piston controller
l ................ stroke length of the piston
h(t) . . . . . . . . . . . . exit gauge
ĥ(t) . . . . . . . . . . . . estimate of exit gauge
href (t) . . . . . . . . . exit gauge reference signal
H(t) . . . . . . . . . . . input gauge
m1 . . . . . . . . . . . . . mass of hydraulic piston
m2 . . . . . . . . . . . . . mass of backup roll
m3 . . . . . . . . . . . . . mass of work roll
M .............. mill modulus (spring constant)
M̂ . . . . . . . . . . . . . . estimated mill modulus
p(t) . . . . . . . . . . . . hydraulic pressure
qf (t) . . . . . . . . . . . fluid flow to hydraulic capsule
v(t) . . . . . . . . . . . . unknown input
v̂(t) . . . . . . . . . . . . unknown input estimate
vc , v1 , v2 . . . . . . . . auxiliary coefficients
y(t) . . . . . . . . . . . . system output
z(t) . . . . . . . . . . . . position of the hydraulic piston

165
 7. Potential applications

zref (t) . . . . . . . . . hydraulic piston position reference signal

µc . . . . . . . . . . . . . . Coulomb friction level
µs . . . . . . . . . . . . . . Stribeck friction coefficient
µv . . . . . . . . . . . . . . viscous friction coefficient

Preliminary reading: Sections 3.2, 3.3, and 3.5

7.1 Introduction
Algorithms developed in Chapter 3 are evaluated using two practical examples. In
Section 7.2 the PE-UIO (see Algorithm 3.1) is used to improve control performance
of a steel rolling mill. Furthermore, the two stage PE-UIO (Algorithm 3.4) is used to
estimate the concentration of river pollutant. Conclusions and a critical appraisal of
practical use of the developed algorithms are presented in Section 7.4.

7.2 Steel rolling mill

In this Section Algorithm 3.1 is applied to a model of a single stand hot strip finishing
rolling mill. Rolling is a process of shaping a metal piece by a reduction of its thickness,
i.e. gauge. The metal is compressed by being progressed between rollers rotating with
the same velocity but in the opposite direction. The final stage of the rolling process is
the finishing mill, where the main goal is to maintain the exit gauge, i.e. the thickness
of the steel strip emerging from the mill, within increasingly tight specifications. The
finishing mill consists of several stands, which consecutively reduce the thickness of the
steel strip. Each of the finishing mill stands is controlled separately.
For the purpose of control, the force acting on the strip (further denoted as ‘roll
force’) is required to be inferred via measurement of a hydraulic actuator force. The
roll force is obtained via measurements of the fluid pressure in a hydraulic actuator
mounted on the top of the stand. However, a problem, which occasionally arises is an
oscillation of the exit gauge. It is believed that this is due to limit cycles, which result
from the nonlinearities in the plant, caused mainly by friction. The contacting surfaces
of the mill are many and the friction forces acting on the elements can be large hence
their impact on the plant behaviour is significant. The friction force is a strongly non-
linear phenomenon, difficult to model and dependent on many operating environment
conditions, such as temperature, properties of lubricant, wear of surfaces, etc., see, for
example, (Papadopoulos & Chasparis 2002, Putra 2004). The non-stationarity of these
conditions cause the friction to be difficult to parametrise.
Algorithm 3.1 is applied to the plant in order to estimate the value of the force
measurement error, which is assumed to be mainly caused by the friction and constitutes
on the unknown input. Based on the estimated value of the error, a correction is made
to the measured value of the force, which is then fed back to the plant.

166
 7. Potential applications

piston position z(t)

hydraulic force Fh (t)

roll force Fr (t)

H(t)
h(t)

work roll

backup roll

Figure 7.1: Schematic illustration of controlled plant

A PE-based friction compensator applied to a deterministic model of a rolling mill

has been previously investigated by the author in (Sumislawska, Burnham, Hearns,
Larkowski & Reeve 2010). Furthermore, in (Sumislawska, Larkowski & Burnham
2010a) an additive measurement noise on the piston position and hydraulic force has
been considered.

7.2.1 Description of the plant

A white-box model of the plant has been originally developed in (Sumislawska 2009).
Furthermore, it has been reported in (Sumislawska, Reeve, Burnham, Pozniak-Koszalka
& Hearns 2009) and (Sumislawska, Burnham, Hearns, Larkowski & Reeve 2010). A
schematic illustration of the plant is presented in Fig. 7.1. The steel strip remains
constantly in a contact with a pair of working rolls, which are supported by the backup
rolls. The entry gauge and the exit gauge of the strip are denoted as H(t) and h(t),
respectively. The hydraulic actuator mounted at the top of the stack changes the
position of the backup, hence the work rolls, and ultimately controls the exit gauge of
the strip (Yildiz, Forbes, Huang, Zhang, Wang, Vaculik & Dudzic 2009).

Control scheme

Harsh temperature conditions close to the rolling mill stand render direct measurement
of the exit gauge impossible (Yildiz et al. 2009), hence a need arises to estimate the
exit gauge from the measured value of the roll force and the mill modulus, i.e. mill
sensitivity to force, see (Yildiz et al. 2009). Due to the fact that the roll force is
inaccessible for measurements, the force in the hydraulic actuator capsule is measured.
Therefore the exit gauge is estimated via:

Fh (t)
C
ĥ(t) = z(t) + (7.1)
M̂

167
 7. Potential applications

where ĥ(t) denotes the estimated exit gauge, M̂ is the estimated value of the mill
modulus M (i.e. mill sensitivity to force), z(t) denotes the hydraulic piston position
and Fh (t) corresponds to the value of the force measured in the hydraulic actuator
capsule, noting that Fh (t) ≈ Fr (t), where Fr (t) denotes the roll force. In fact it is
the difference between the measured Fh (t) and the actual Fr (t) which constitutes the
unknown input. The actual exit gauge, denoted h(t), is is given by:

Fr (t)
C
h(t) = z(t) + (7.2)
M

In order to improve the robustness of the control loop a compensation variable C < 1
is introduced.
The control scheme of the rolling mill is presented in Fig. 7.2. The controller gains
G1 and G2 are given by:

G1 = 1 + (1 − C)
k3 M̂ + k3
, G2 = (7.3)
M̂ M̂ + k3 + Ck3

where k3 denotes the steel strip sensitivity to force (strip modulus). The term Kp
denotes the proportional gain of the hydraulic piston position controller and defines
a relation between the piston position error and the fluid flow, denoted qf (t), to the
hydraulic actuator capsule.
href (t) + + zref (t) + qf (t) Fh (t) h(t)
G1 G2 Kp
_ _ Actuator Stack z(t)
+

C
M̂

+
ĥ(t) +

Figure 7.2: Control loop, href (t) – exit gauge reference signal, zref (t) – piston po-
sition reference signal, Fh (t) – hydraulic force, qf (t) – flow of hydraulic
fluid

7.2.2 Plant model

Stack model

The stack of rolls (i.e. backup and work rolls with a steel strip between them, further
referred to as the stack) is modelled by making use of a classical mass-spring-damper
model representation, see Fig. 7.3. Due to the symmetrical construction of the stack,
only the upper backup and work rolls are taken into consideration. The values of the
model parameters are given in Table 7.1. In the further analysis the damper denoted

168
 7. Potential applications

d1 is replaced by a friction model, which introduces a nonlinear dependency between

the piston velocity and the friction force.
Fh (t)

m1
z(t)
d1
k1

m2

d2 k2

m3
h(t)

d3
k3

Figure 7.3: Mass-spring-damper representation of the stack, m1 - hydraulic piston

mass, m2 - backup roll mass, m3 - work roll mass, k1 - spring coefficient
between piston and backup roll, k2 - spring coefficient between backup
and work roll, k3 - spring coefficient of the strip

Parameter Value Unit

m1 1 ⋅ 102 kg
m2 5 ⋅ 104 kg
3
m3 6 ⋅ 10 kg
N
k1 1 ⋅ 1010 m
N
k2 1 ⋅ 1010 m
N
k3 3 ⋅ 109 m
kg
d1 1 ⋅ 107 s
kg
d2 5 ⋅ 106 s
kg
d3 5 ⋅ 106 s
N
M 5 ⋅ 109 m

Table 7.1: Parameters of rolling mill model

Friction model

The friction is modelled as a sum of three components: Coulomb friction, viscous

friction and Stribeck friction. The latter is also named stiction (Putra 2004):

Ff ric (t) = Fc (t) + Fv (t) + Fs (t) (7.4)

169
 7. Potential applications

where the term Ff ric (t) denotes the total frictional force, whilst Fc (t), Fv (t) and Fs (t)
refer to the Coulomb, viscous and Stribeck friction, respectively. The Coulomb friction
is modelled as:
Fc (t) = −µc sign(ż(t))(1 − e∣ )
ż(t)
∣
vc (7.5)

where the term µc denotes the Coulomb friction level, whilst the exponential term is
introduced in order to avoid a zero-crossing discontinuity. The element related to the
viscous friction is modelled as a linear function of the hydraulic piston velocity, denoted
ż(t):
Fv (t) = −µv ż(t) (7.6)

where the term µv denotes the viscous damping of the frictional force. The Stribeck
friction (stiction) model is given by:

Fs (t) = −µs sign(ż(t))(1 − e )e

ż(t) ż(t)
∣ ∣ ∣ ∣
v1 v2 (7.7)

where the term µs determines the magnitude of the static friction, whilst v1 and v2 are
utilised to shape the stiction model.

Hydraulic servo system model

Dependency between the fluid flow, denoted qf (t), into the capsule and the pressure
denoted p(t) acting on the piston area denoted Ap is represented by the following linear
relation, see e.g. (Jelali & Kroll 2003):

∫ qf (t)dt − Ap l
p(t) = Kc (7.8)
Ap l

where l denotes the stroke length of the piston, and the term Kc corresponds to the
hydraulic oil compressibility coefficient. Therefore, the force denoted Fh (t) acting on
the hydraulic piston is given by:

Fh (t) = Ap p(t) (7.9)

The values of the hydraulic actuator model parameters are given in Table 7.2.

Parameter Value Unit

Ap 0.331 m2
l 0.1 m
N
Kp 7.0 m
Kc 3.32 ⋅ 109 Pa

Table 7.2: Parameters of hydraulic actuator model

170
 7. Potential applications

7.2.3 Simulation results

To simulate the plant the full nonlinear model is used, however to generate the PE-
UIO a linear model is required. The nonlinear model of the system has been linearised
by replacing the damping coefficient d1 (i.e. the ratio of the friction force to the
piston velocity) by its nominal value equal to 107 kg
s . The white-box model described
in Subsection 7.2.2 is discretised using a sampling interval of 10 ms, which corresponds
to the sampling interval of the controller. The minimal state-space representation of
the model is a one-input two-output 7th order system, which is conveniently obtained
using the Matlab linmod function.
The resulting discretised form of the linearised closed-loop system has then been
used for friction force estimator design. The reference signal is considered to be the
input to the system. The two measured outputs of the system are the roll force (i.e.
the force acting on the strip) and the piston position, given by, respectively, the first
and the second elements of the output vector y(t), see (7.10). It is convenient to scale
C
the roll force by to ensure numerical scalability.
M̂

y(t) = [ F (t) z(t) ]

T
C
(7.10)
M̂ r

Recall that due to the fact that the roll force Fr (t) is inaccessible, the hydraulic force
Fh (t) is measured. Subsequently, the force measurement is affected by the friction force
and the parasitic dynamics of the stack. Due to the fact that the bandwidth of the
parasitic dynamics of the stack exceeds the sampling frequency, it can be assumed that
the frictional force has the most significant contribution to the roll force measurement
error. Therefore, the difference between the roll force Fr (t) and the force measured
in the hydraulic capsule Fh (t) is considered to be the unknown input to the system,
further referred to as v(t), where v(t) = C
(Fr (t) − Fh (t)). (The factor C
is used to
M̂ M̂
ensure to ensure numerical scalability.) Hence, the matrices G and H of the system
(3.1) are:
H =[ 1 0 ]
T
G = 0, (7.11)

Subsequently, the exit gauge change is estimated via:

Fh (t) − v̂(t)
C
ĥ(t) = z(t) + (7.12)
M̂

where v̂(t) is the correction term (i.e. the estimated force the measurement error v(t),
corresponding to the unknown input obtained from the PE-UIO).
Engineering knowledge and past experience of technicians with the plant indicate
that it is reasonable to assume that the piston position and hydraulic force measure-
ments are affected by white, zero-mean, Gaussian, mutually uncorrelated noise se-
quences, whose standard deviations are, respectively, 0.1µm and 1000N. The PE-UIO
algorithm with s = 4 samples is used to obtain v̂(t).

171
 7. Potential applications

0.03
no compensation
with compensation
0.02
Gauge error [mm]

0.01

−0.01

−0.02

−0.03
10 11 12 13 14 15 16 17 18 19 20
Time [s]

Figure 7.4: Friction compensation effect on the exit gauge error

5
x 10
2
estimated
1.5 true

1
Friction force [N]

0.5

−0.5

−1

−1.5

−2
15 15.5 16 16.5 17 17.5 18 18.5 19 19.5 20
Time [s]

Figure 7.5: Friction force estimation

Fig. 7.4 presents the simulated results for the cases of no compensation and with the
unknown input compensation applied. The grey dashed line corresponds to a simulated
reconstruction of a typical limit cycle condition found in practice. The black solid line
corresponds to the compensated case and clearly indicates that the limit cycle ampli-
tude is significantly reduced, implying potential for improved product quality. Fig. 7.5
shows the actual unknown input and the estimated unknown input corresponding to
the simulated condition in Fig. 7.4. The PE-UIO accurately estimates the friction
force affecting the exit gauge and a subsequent feedback compensation that utilises the
estimated unknown input results in a significant improvement in control.

7.3 Hydrological application

The second example is based on the data collected during a potassium bromide (KBr)
tracer experiment carried out in a wetland area by Martinez & Wise (2003). A

172
 7. Potential applications

schematic illustration of the experiment is depicted in Fig. 7.6. A tracer has been
poured into the river at the point (1). Two tracer concentration sensors have been
placed downstream, at points (2) and (3), whose readings are denoted, respectively,
as v(t) and y(t). A linear model of the relation between v(t) and y(t), where v(t) is
the input to the system, whilst y(t) is the output, has been developed in (Young &
Sumislawska 2012). The input v(t), further referred to as upstream tracer concentra-

KBr v(t) y(t)

(1) (2) (3)

flow direction

Figure 7.6: Tracer experiment

tion, and the output y(t), the tracer concentration measured downstream, are plotted
as the grey line and black line, respectively, in Fig. 7.7. The aim of this simulation is
to use the two stage PE-UIO (Algorithm 3.4) scheme to estimate the input v(t) based
on output measurements and the knowledge of the system model. The system can be

5
v(t)
y(t)
v(t) and y(t) [mg/l]

0
0 100 200 300 400 500 600 700 800 900
Time [hours]

Figure 7.7: Input and output signals in tracer experiment

approximated by a linear second order model, with time constants of 17.4 and 83.7
hours, i.e. a stiff system. The discrete time model of the two-hourly sampled system is
given by (Young & Sumislawska 2012):

Gv (z) =
0.017591(z + 4.302)(z − 0.9735)
(z − 0.9764)(z − 0.8916)
(7.13)

Note that the same model has been used in Example 3.7 to obtain the unknown in-
put of the simulated system (in contrast to this example, where real input and output
measurements are used). The model is nonminimum-phase. Although the standard
PE-UIO can cope with the zero at −4.302, the zero at 0.9735 requires the two stage

173
 7. Potential applications

PE-UIO to be used. For the design of the unknown input reconstruction it has been
assumed that the measurements are affected by white, zero-mean, Gaussian, mutually
uncorrelated measurement noise. The parity space order has been set to 15 samples,
which leads to an estimation time lag of 7 samples. The result of an unknown input
estimation using the two stage PE-UIO is compared with the input reconstructed us-
ing the INPEST, see Fig. 7.8. The parameters of the INPEST method are τ̊ = 7 and
q̊e = 0.8 obtained for λ = 0.001, see (Young & Sumislawska 2012). It is noted that both

Error
0.5
0
−0.5
−1
0 50 100 150 200 250 300 350 400
Unknown input estimate [mg/l]
5
true
PE−UIO
4 INPEST

0
0 50 100 150 200 250 300 350 400
Time [samples]

Figure 7.8: Result of unknown input estimation

methods give similar results. Both methods detect a rise of v(t) in approximately the
33th sample, whereas the measured input starts rising at approximately the 42th sam-
ple. Furthermore, after the 200th sample the estimation errors of the PE-UIO and the
INPEST are virtually the same. It is believed that the input reconstruction discrep-
ancies are caused mainly by the modelling inaccuracy, presumably caused by system
nonlinearities. This hypothesis is supported by Fig. 7.9, which compares the measured
output with the model output. The simulated output starts rising approximately 10
samples after the rise of the measured output. Furthermore, the observed characteris-
tic ‘bumps’ of the measured output between 100 and 150 samples result in the input
reconstruction error pattern visible in the upper subfigure of Fig. 7.8.

7.4 Critical appraisal of practical application of developed

methods
A simulation study of a single finishing stand of a steel rolling mill has demonstrated
promising results. The force measurement error has been estimated quite accurately,

174
 7. Potential applications

4
measured
Dye concentration [mg/l] simulated
3

0
0 50 100 150 200 250 300 350 400 450
Time [samples]

Figure 7.9: Model output vs. measured output

and the compensation significantly reduced the amplitude of limit cycles. The high
frequency and low amplitude oscillations, which may be observed after enhancement of
the control, are probably the result of unknown input estimation delay. In the industrial
plant, however, unmeasured variations of the input gauge H occur, which should be
treated as a disturbance. Thus, before application to an actual plant, the possibility of
disturbance decoupling in the PE-UIO needs to be explored.
The two stage PE-UIO has estimated the tracer concentration in the river accu-
rately. Further improvement could possibly be achieved, if, instead of the assumption
of a white measurement noise, a coloured process noise model, which can explain dis-
crepancy between modelled and simulated output (model mismatch), is assumed.

175
Chapter 8

Conclusions & further work

8.1 Conclusions
This thesis presents novel developments in the fields of unknown input reconstruction
and fault detection, isolation and identification. The developed algorithms are applica-
ble to time-invariant, discrete-time systems. Most of the research is devoted to linear
systems, except for the unknown input reconstruction method presented in Chapter 4,
which has been designed for a class of nonlinear systems, namely, Hammerstein-Wiener
systems. Two potential applications for the algorithms developed in Chapter 3 have
been proposed and promising results demonstrated via simulation studies.
This section is divided into three logical parts. Subsection 8.1.1 summarises the
development of unknown input reconstruction schemes. In Subsection 8.1.2 fault de-
tection and diagnosis algorithms are concluded. The main contributions of this thesis
are summarised in Subsection 8.1.3.

8.1.1 Unknown input reconstruction

In Chapter 3 a novel scheme combining PE and the Lagrange multiplier optimisation
method for unknown input reconstruction of MIMO stochastic systems has been de-
vised. Due to that fact that the PE-UIO utilises parity equations, i.e. both input and
output signals are filtered in an analogous manner, the method is suitable for systems
in the EIV framework. It is assumed that the system input is affected by white, Gaus-
sian, zero-mean measurement noise, whilst the output is subjected to coloured noise,
which may represent a combination of measurement and process noise sequences. In
particular, the methods can be applied to systems with well known noise models, such
as ARX, ARMAX or OE. The PE-UIO requires the knowledge of the system model
and, if the input is subjected to measurement noise (EIV framework), at least the ratio
between the variances of the input and output noise sequences. Otherwise, if the input
can be measured directly (i.e. there is no noise affecting the input), the knowledge of
the noise variance is not required to be known explicitly for the design of the input

176
 8. Conclusions & further work

reconstruction filter (it is, however, required to know the noise model). In the case of
OE systems the design procedure can be simplified, which has also been demonstrated.
The only tuning parameter for the PE-UIO is the parity space order. By increasing it,
the bandwidth of the filter is reduced, and, consequently, noise filtering properties are
improved. On the other hand, reduced bandwidth causes an estimation lag. Thus, the
trade-off between noise filtering and estimation lag as well as an a’priori knowledge of
the bandwidth of the reconstructed signal needs to be taken into consideration. The
algorithm is applicable to both minimum-phase and nonminimum-phase systems.
The drawback of the PE-UIO is that it may produce a distorted unknown input
estimate, if a zero of the system transfer function to an unknown input is unity (a system
with a derivative term) or lies close to unity. To tackle this problem an extension to
the PE-UIO, i.e. a two stage PE-UIO, has been proposed. The two stage PE-UIO is
applicable to systems, whose minimum-phase zeros lie close to unity or its zeros are
equal unity. It has been demonstrated that the two stage algorithm has superior noise
filtering properties compared to the standard PE-UIO, however, it may introduce larger
estimation lag. Both, the standard and the two stage input reconstruction algorithms,
are computationally simple. The filter parameters need to be calculated only once
before the filter is applied to the system.
Both, the standard (single stage) PE-UIO and the two stage PE-UIO, have been
compared with two other methods found in the literature: a Kalman filter-based MVU
and the INPEST method, based on linear quadratic control. A simulation study has
revealed superior noise filtering properties of the PE-UIO compared to the MVU. This
is due to the adjustable bandwidth of the PE-UIO (which, however, causes the trade-
off between the noise filtering and estimation delay). Furthermore, the MVU in the
case of a single output system resembles a naive inversion, thus it cannot be used
for unknown input reconstruction of single output nonminimum-phase systems. The
INPEST method has shown comparable results to those of the PE-UIO (both single
stage and the two stage).
Potential industrial applications of the proposed unknown input reconstruction
schemes have been demonstrated via simulation studies in Chapter 7. The PE-UIO
has been used to improve the control performance of a steel rolling mill, by recon-
struction of a parasitic friction force. The two stage PE-UIO has been proposed in a
hydrological application in order to estimate the level of pollutant in a river.
In Chapter 4 the PE-UIO has been extended to Hammerstein-Wiener systems, i.e.
systems which can be modelled as a linear dynamic block preceded and followed by
a static nonlinearity. The algorithm has been developed for a system with a single
measurable input, single output, and a single unknown input to be reconstructed in
an EIV framework, where both measured input and output are subjected to white,
Gaussian, zero-mean mutually uncorrelated noise sequences. As the system operating
point changes, the influence of the input and output noise on the unknown input es-

177
 8. Conclusions & further work

timate varies. Thus, the algorithm needs to adapt to these changes. Two versions of
the scheme are proposed. In the first version the parity space order remains constant,
whilst the filter parameters vary at each time sample. In the second version the parity
space order varies according to the system operating point. Furthermore, assuming a
Gaussian distribution of the measurement noise, a method for computation of the confi-
dence bounds has been proposed. The simulation study has demonstrated applicability
of proposed algorithms to the particular class of nonlinear systems. It has also been
shown that for relatively mild nonlinearities a linear algorithm can be used in order to
reduce the computational effort.

8.1.2 Fault detection and diagnosis

A robust fault detection filter of (Chen & Patton 1999) has been extended to system
with unstable invariant zeros, which improved the applicability of the scheme. Fur-
thermore, a modification has been proposed which allowed relaxation of a strict rank
condition. The robust fault isolation filter proposed in Chapter 5 has an equivalent
eigenstructure to the fault isolation filter proposed in (Chen & Speyer 2006a), how-
ever, it has been demonstrated that the design procedure presented in Chapter 5 is
computationally simpler. It has been shown that the devised robust fault detection
filter is completely decoupled from disturbances. The user can influence the transient
behaviour of the residual response to faults via assignment of poles of the filter.
Furthermore, by building on the finite time convergent state observer of Engel &
Kreisselmeier (2002), the proposed robust fault detection filter has been used to design
robust PE of an arbitrary order. Analogously, left eigenstructure assignment has been
used to design the PE of any user defined order. It has been demonstrated that both
algorithms provide complete disturbance decoupling.
In Chapter 6 the robust PE via right eigenstructure assignment have been extended
to the fault isolation and identification filter case. Decoupling properties of the robust
PE presented in Chapter 5 have been used to design a directional residual set. Fur-
thermore, by application of a similarity transformation and by setting to unity the
steady state gain of the residual response of the filter to a given fault, a filter has been
obtained, which provides an estimate of the fault vector, i.e. reconstructs the fault
vector. In the case when the number of outputs exceeds the number of possible faults,
the design freedom has been used to devise a control variable, which remains zero when
any of the modelled faults occurs. Deviation of the control variable from zero indicates
an occurrence of a fault which has not been covered by the model. Applicability of
the novel PE design for fault identification to stochastic systems in the EIV framework
has been considered and a simple method for calculation of thresholds, whose violation
indicates a fault, has been proposed. Efficacy of the method has been demonstrated
using a numerical example.

178
 8. Conclusions & further work

8.1.3 Contributions
The main contributions of this thesis are briefly summarised in order of importance as
follows:

1. Incorporation of the Lagrange multiplier optimisation method into PE design in

order to minimise the noise effect on the unknown input estimate (Chapter 3).

2. Extension of the proposed unknown input estimator to Hammersten-Wiener sys-

tems (Chapter 4).

3. Application of the novel PE-based unknown input reconstruction method for en-
hancement of a control loop in a single stand of a rolling mill and for a hydrological
application (Chapter 7).

4. Use of right and left eigenstructure assignment to develop robust fault detection
PE of user defined order (Chapter 5).

5. Extension of the robust fault detection filter via right eigenstructure assignment
to systems with unstable invariant zeros (Chapter 5).

6. Use of right eigenstructure assignment to develop PE of arbitrary order for the

purpose of fault isolation and identification (Chapter 6).

8.2 Further work

The following aspects are considered as further work regarding the PE-UIO for linear
stochastic systems:

● Up to date the PE-UIO can be applied to systems where a single unknown input
needs to be reconstructed. Thus, an extension of the algorithm to systems with
multiple unmeasurable inputs could be considered.

● Application of a disturbance decoupling scheme to the PE-UIO could improve

applicability of the algorithm.

● The single stage PE-UIO produces a distorted unknown input estimate when a
zero of the system response to an unknown input lies close to unity. The two
stage PE-UIO copes with such a situation if the problematic zero lies inside the
unit circle. However, the problem remains open for the cases when a system
nonminimum-phase zero lies close to unity.

The following aspects of the input reconstruction scheme for Hammerstein-Wiener sys-
tems could be taken into consideration:

● Extension of the algorithm to the multivariable case, with multiple measured

inputs and outputs as well as multiple unknown inputs to be reconstructed.

179
 8. Conclusions & further work

● Algorithms presented in Chapter 4 require calculation of filter parameters at

each time sample, thus they need a computational power which is not negligible.
Consequently, an optimisation of the procedure is considered as an interesting
aspect of future work.

● The scheme utilises a direct inversion of the output transforming nonlinearity,

hence it is not applicable to systems where the output nonlinearity is not in-
vertible. It would be interesting to consider e.g. adaptive learning methods to
reconstruct the input to the noninvertible block and hence improve the applica-
bility of the method.

Additional research on the topic of fault detection and diagnosis could include:

● Robust fault detection for stochastic systems. Design freedom of both robust
PE and a robust asymptotic filter could be used to minimise the influence of the
noise of the residual. Also calculation of thresholds, whose violation indicates the
presence of faults has not been discussed for the robust fault detection filter (it has
been discussed only for the fault isolation and identification filter in Chapter 6).

● Exploring applicability of the robust fault detection scheme to systems where the
disturbance direction vectors ei combine with each other to create new invariant
zeros.

● Extension the fault isolation and identification scheme to multidimensional faults.

Furthermore, the following future research directions could also be considered:

● Extension of the proposed schemes to time-varying and uncertain systems.

● Extension of the proposed robust fault detection and diagnosis methods to non-
linear systems; particularly, Hammerstein-Wiener and bilinear systems could be
considered.

● Whilst two of the developed algorithms have been applied to practical applica-
tions, it would be desirable to evaluate the other methods developed within this
thesis to real world applications to assess their potential benefits.

180
References

Alexandru, M. (2003), ‘Analysis of induction motor fault diagnosis with fuzzy neural
network’, Applied Artificial Intelligence: An International Journal 17(2), 105–133.

Alexandru, M. & Popescu, D. (2004), Neuro-fuzzy diagnosis in final control elements

of AC motors, in ‘Proceedings of the 2004 American Control Conference’, Boston,
MA, USA.

Anbumani, K., Patnaik, L. & Sarma, I. (1981), ‘Self-tuning minimum-variance control

of nonlinear systems of the Hammerstein model’, IEEE Transactions on Automatic
Control 26(4), 959–961.

Antsaklis, P. (1978), ‘Stable proper nth-order inverses’, IEEE Transactions on Auto-

matic Control 23(6), 1104–1106.

Ashari, A., Sedigh, A. & Yazdanpanah, M. (2005a), Output feedback reconfigurable

controller design using eigenstructure assignment: post fault order change, in ‘Pro-
ceedings of the International Conference on Control and Automation, 2005. ICCA
’05’, Vol. 1, Budapest, Hungary, pp. 474–479.

Ashari, A., Sedigh, A. & Yazdanpanah, M. (2005b), ‘Reconfigurable control system

design using eigenstructure assignment: static, dynamic and robust approaches’,
International Journal of Control 78(13), 1005–1016.

Basile, G. & Marro, G. (1969), ‘Controlled and conditioned invariant subspaces in linear
system theory’, Journal of Optimisation Theory and Applications 33(5), 305–315.

Basile, G. & Marro, G. (2002), Controlled and Conditioned Invariants in Linear Sys-
tems Theory, Department of Electronics, Systems and Computer Science, Univer-
sity of Bologna, Italy.

Basile, G. & Marro, G. (2010), ‘The geometric approach toolbox’, http://www3.deis.

unibo.it/Staff/FullProf/GiovanniMarro/geometric.htm\#reftools.

Basilevsky, A. (1983), Applied Matrix Algebra in the Statistical Sciences, North-

Holland, New York/Amsterdam/Oxford.

181
 REFERENCES

Beard, R. V. (1971), Failure Accommodation in Linear Systems through Self-

reorganisation, PhD thesis, Massachusetts Institute of Technology, USA.

Berriri, H., Naouar, M. W. & Slama-Belkhodja, I. (2011), Parity space approach for
current sensor fault detection and isolation in electrical systems, in ‘Proceedings
of the 8th International Multi-Conference on Systems, Signals & Devices’, Sousse,
Tunisia.

Bertsekas, D. P. (1982), Constrained Optimisation and Lagrange Multiplier Methods,

Academic press, Inc., London.

Bhattacharjee, A., Sengupta, A. & Sutradhar, A. (2010), Nonparametric modeling

of glucose-insulin process in IDDM patient using Hammerstein-Wiener model, in
‘Proceedings of 11th International Conference on Control Automation Robotics &
Vision (ICARCV),’, Singapore.

Bhattacharjee, N. & Roy, B. (2010), Fault detection and isolation of a two non-
interacting tanks system using partial PCA, in ‘Proceedings of 2010 International
Conference on Industrial Electronics, Control & Robotics (IECR)’, Rourkela, In-
dia.

Campos-Delgado, D. (2011), ‘An observer-based diagnosis scheme for single and si-
multaneous open-switch faults in induction motor drives’, IEEE Transactions on
Industrial Electronics 58(2), 671–679.

Celka, P. & Colditz, P. (2002), ‘Nonlinear nonstationary Wiener model of infant EEG
seizures’, IEEE Transactions on Biomedical Engineering 49(6), 556–564.

Chabir, K., Sauter, D. & Keller, J. (2009), Design of fault isolation filter under network
induced delay, in ‘Proceedings of 2009 IEEE Conference on Control Applications,
(CCA) & Intelligent Control, (ISIC)’.

Chan, C., Hua, S. & Hong-Yue, Z. (2006), ‘Application of fully decoupled parity equa-
tion in fault detection and identification of dc motors’, IEEE Transactions on
Industrial Electronics 53(4), 1277–1284.

Chen, B. & Nagarajaiah, S. (2007), ‘Linear-matrix-inequality-based robust fault detec-

tion and isolation using the eigenstructure assignment method’, Journal of Guid-
ance, Control, and Dynamics 30(6), 1831–1835.

Chen, J. & Patton, R. J. (1999), Robust Model-Based Fault Diagnosis for Dynamic
Systems, Kluver Academic Publishers, Boston.

Chen, R. H. & Speyer, J. L. (2006a), Detection filter analysis and design using
eigenstructure assignment, in ‘Proceedings of the American Control Conference
(ACC’06)’, Minneapolis, Minnesota USA.

182
 REFERENCES

Chen, R. & Speyer, J. (2006b), Generalization of the detection filter using spectral the-
ory, in ‘Proceedings of the American Control Conference (ACC’06)’, Minneapolis,
Minnesota USA.

Chen, R. & Speyer, J. (2007), Spectral analysis and design of detection filter for
multiple-dimensional faults, in ‘Proceedings of the American Control Conference
(ACC’07)’, New York City, USA.

Chow, E. & Willsky, A. (1984), ‘Analytical redundancy and the design of robust failure
detection systems’, IEEE Transactions on Automatic Control 29(7), 603–614.

Craddock, R. V. (1962), ‘Take-off monitoring apparatus for aircraft’, United States

Patent, No. 3,034,096, 8 May.

Crama, P. & Rolain, Y. (2002), Broadband measurement and identification of a Wiener-

Hammerstein model for an RF amplifier, in ‘Proceedings of 60th ARFTG Confer-
ence Digest’.

Crama, P. & Schoukens, J. (2004), ‘Hammerstein-Wiener system estimator initializa-

tion’, Automatica 40(9), 1543–1551.

Czop, P. (2011), ‘Application of inverse data-driven parametric models in the recon-

struction of passenger vehicle wheel vertical movement under ride conditions’,
Journal of Vibration and Control 18(8), 1133–1140.

Darouach, M. & Zasadzinski, M. (1997), ‘Unbiased minimum variance estimation for

systems with unknown exogenous inputs’, Automatica 33(4), 717–719.

De-Feng, H., Li, Y. & Guo-Shi, Y. (2010), A pole placement-based NMPC algorithm
of constrained Hammerstein systems, in ‘Proceedings of 29th Chinese Control
Conference (CCC 2010)’, Beijing, China.

Dever, J. A. (1960), ‘Control apparatus’, United States Patent, No. 2,931,763, 5 May.

Ding, S. X. (2008), Model-based Fault Diagnosis Techniques: Design Schemes, Algo-

rithms and Tools, Springer Verlag, Berlin.

Ding, S. X., Zhong, M., Bingyong, T. & Zhang, P. (2001), An LMI approach to the
design of fault detection filter for time-delay LTI systems with unknown inputs,
in ‘Proceedings of the American Control Conference (ACC’01)’, Arlington, VA,
USA.

Dobrowiecki, T. & Schoukens, J. (2002), Cascading Wiener-Hammerstein systems, in

‘Proceedings of the IEEE Instrumentation and Measurement Technology Confer-
ence’, Anchorage, AK, USA.

183
 REFERENCES

Dorato, P. (1969), ‘On the inverse of linear dynamical systems’, IEEE Transactions on
Systems Science and Cybernetics 5(1), 43–48.

Douglas, R. K., Speyer, J. L., Mingori, D. L., Chen, R. H., Malladi, D. P. & Chung,
W. H. (1996), ‘Fault detection and identification with application to advanced ve-
hicle control systems: Final report’, Research Reports, California Partners for Ad-
vanced Transit and Highways (PATH), Institute of Transportation Studies (UCB),
UC Berkeley.

Douglas, R. & Speyer, J. (1995), Robust fault detection filter design, in ‘Proceedings of
the American Control Conference (ACC’95)’, Vol. 1, Seattle, WA, USA, pp. 91–96.

Duan, G. & Patton, R. J. (2001), ‘Robust fault detection using Luenberger-type un-
known input observers – a parametric approach’, International Journal of Systems
Science 32(4), 533–540.

Dutka, A., Javaherian, H. & Grimble, M. J. (2009), Model-based engine fault detection
and isolation, in ‘Proceedings of the American Control Conference (ACC’09)’, St.
Louis, MO, USA.

Edelmayer, A. (2005), Fault Detection in Dynamic Systems: From State Estimation

to Direct Input Reconstruction Methods, D.Sc. thesis, Hungarian Academy of Sci-
ences, Budapest, Hungary.

El-Ghezawi, O., Billings, S. & Zinober, A. (1983), Variable-structure systems and sys-
tem zeros, in ‘IEE Proceedings Control Theory and Applications’, Vol. 130, pp. 1–
5.

Engel, R. & Kreisselmeier, G. (2002), ‘A continuous-time observer which converges in

finite time’, IEEE Transactions on Automatic Control 47(7), 1202–1204.

Fagarasan, I. & St. Iliescu, S. (2008), Parity equations for fault detection and isolation,
in ‘Proceedings of IEEE International Conference on Automation, Quality and
Testing, Robotics, (AQTR 2008)’, Cluj-Napoca, Romania.

Fantuzzi, C., Simani, S. & Beghelli, S. (2001), Robust fault diagnosis of dynamic pro-
cesses using parametric identification with eigenstructure assignment approach, in
‘Proceedings of the 40th IEEE Conference on Decision and Control’, Orlando, FL,
USA.

Fernando, T. & Trinh, H. (2006), Design of reduced-order state/unknown input ob-

servers: A descriptor system approach, in ‘Proceedings of the 2006 IEEE Interna-
tional Conference on Control Applications’, Munich, Germany.

Floquet, T. & Barbot, J. (2006), ‘State and unknown input estimation for linear
discrete-time systems’, Automatica 42(11), 1883–1889.

184
 REFERENCES

Fragkoulis, D., Roux, G. & Dahhou, B. (2009), A global scheme for multiple and
simultaneous faults in system actuators and sensors, in ‘Proceedings of the 6th In-
ternational Multi-Conference on Systems, Signals and Devices (SSD ’09)’, Djerba,
Tunisia.

Frank, P. M. & Wünnenberg, J. (1989), Fault Diagnosis in Dynamic Systems: Theory

and Application, Prentice Hall, chapter Robust Fault Diagnosis Using Unknown
Input Schemes, pp. 47–98.

Fravolini, M., Brunori, V., Campa, G., Napolitano, M. & La Cava, M. (2009), ‘Struc-
tural analysis approach for the generation of structured residuals for aircraft FDI’,
IEEE Transactions on Aerospace and Electronic Systems 45(4), 1466–1482.

Fruzzetti, K., Palazoglu, A. & McDonald, K. (1997), ‘Nonlinear model predictive con-
trol using Hammerstein models’, Journal of ProcessControl 7(1), 31–44.

Fu, H., Kirtikar, S., Zattoni, E., Palanthandalam-Madapusi, H. & Bernstein, D. (2009),
Approximate input reconstruction for diagnosing aircraft control surfaces, in ‘Pro-
ceedings of the AIAA Guidance, Navigation, and Control Conference’, Chicago,
Illinois.

Fu, H., Yan, J., Santillo, M. A., Palanthandalam-Madapusi, H. & Bernstein, D. (2009),
Fault detection for aircraft control surfaces using approximate input reconstruc-
tion, in ‘Proceedings of the American Control Conference (ACC’09)’, St. Louis,
MO, USA.

Gao, Z., Breikin, T. & Wang, H. (2007), ‘High-gain estimator and fault-tolerant design
with application to a gas turbine dynamic system’, IEEE Transactions on Control
Systems Technology 15(4), 740–753.

Garrick, B. J., Gekler, W. C., Goldfisher, L., Karcher, R. H., Shimizu, B. & Wilson,
J. H. (1967), Reliability analysis of nuclear power plant protective system, Tech-
nical Report HN-190 AEC Research & Development Report, Holmes & Narver,
Inc., Nuclear Division.

Gertler, J. & Kunver, M. (1995), ‘Optimal residual decoupling for fault diagnosis’,
International Journal of Control 61(2), 395–421.

Gertler, J. & Singer, D. (1990), ‘A new structurel framework for parity equation-based
failure detection and isolation’, Automatica 26(2), 381–388.

Ghahremani, E. & Kamwa, I. (2011), Simultaneous state and input estimation of a

synchronous machine using the extended Kalman filter with unknown inputs, in
‘Proceedings of the 2011 IEEE International Electric Machines & Drives Confer-
ence (IEMDC)’, pp. 1468– 473.

185
 REFERENCES

Gillijns, S. & De Moor, B. (2007a), ‘Unbiased minimum variance input and state esti-
mation for linear discrete-time systems’, Automatica 43(1), 111–116.

Gillijns, S. & De Moor, B. (2007b), ‘Unbiased minimum variance input and state es-
timation for linear discrete-time systems with direct feedthrough’, Automatica
43(5), 934–937.

Gu, D. & Poon, F. W. (2003), ‘A robust fault-detection approach with application

in a rolling-mill process’, IEEE Transactions on Control Systems Technology
11(3), 408–414.

Halmos, P. R. (1958), Finite-dimensional vector spaces, Springer.

Halton, H. (1963), ‘Design philosophy of an automatic checkout and launch system for
a drone’, IEEE Transactions on Aerospace 1(2), 538–546.

Hou, M. & Müller, P. (1992), ‘Design of observers for linear systems with unknown
inputs’, IEEE Transactions on Automatic Control 37(6), 871–875.

Hsieh, C. (2000), ‘Robust two-stage Kalman filters for systems with unknown inputs’,
IEEE Transactions on Automatic Control 45(12), 2374–2378.

Ibaraki, S., Suryanarayanan, S. & Tomizuka, M. (2005), ‘Design of Luenberger state

observers using fixed-structure H∞ optimization and its application to fault de-
tection in lane-keeping control of automated vehicles’, IEEE/ASME Transactions
on Mechatronics 10(1), 34–42.

Ibnkahla, M. (2002), ‘Natural gradient learning neural networks for adaptive inversion
of Hammerstein systems’, IEEE Signal Processing Letters 9(10), 315–317.

Isermann, R. (2005), ‘Model-based fault-detection and diagnosis–status and applica-

tions’, Annual Reviews in Control 29, 71–85.

Isermann, R. & Balle, P. (1997), ‘Trends in the application of model-based fault detec-
tion and diagnosis of technical processes’, Control Engineering Practice 5(5), 709–
719.

Janis, J. P. (1963), ‘Checkout methods for space vehicle subsystems’, IEEE Transac-
tions on Aerospace 1(2), 547–549.

Jelali, M. & Kroll, A. (2003), Hydraulic servo-systems: modelling, identification and

control, Springer-Verlag, London.

Jirauch, D. H. (1967), ‘Software design techniques for automatic checkout’, IEEE Trans-
action on Aerospace and Electronic Systems AES-3(6), 934–940.

Jones, H. L. (1973), Failure Detection in Linear Systems, PhD thesis, Massachusetts

Institute of Technology, USA.

186
 REFERENCES

Kaufmann, R. H. & Finison, H. J. (1952), D-C Power Systems for Aircrafts, John
Wiley & Sons, Inc., New York.

Keller, J. & Sauter, D. (2010), A variable geometric state filtering for stochastic linear
systems subject to intermittent unknown inputs, in ‘Proceedings of the Conference
on Control and Fault-Tolerant Systems (SysTol)’, pp. 558–563.

Kennedy, J. J. (1970), ‘Fault detection monitor circuit provides ”self-heal capability”

in electronic modules: A concept’, NASA Tech. Brief 70-10515.

Kirtikar, S., Palanthandalam-Madapusi, H., Zattoni, E. & Bernstein, D. S. (2009),

l -delay input recontruction for discrete-time linear systems, in ‘Proc. of the Con-
ference on Decision and Control’, Shanghai, China, pp. 1848–1853.

Korbicz, J., Koscielny, J. M., Kowalczuk, Z. & Cholewa, W., eds (2003), Fault Diag-
nosis: Models, Artificial Intelligence, Applications, Springer.

Kowalczuk, Z. & Suchomski, P. (2005), Entirely left eigenstructure-assignment for fault

diagnosis observers, in ‘Proceedings of the 16th IFAC World Congress, 2005’,
Vol. 16.

Lajic, Z., Blanke, M. & Nielsen, U. D. (2009), Fault detection for shipboard monitor-
ing Volterra kernel and Hammerstein model approaches, in ‘Proceedings of IFAC
Symposium on Fault Detection, Supervision and Safety of Technical Processes’,
pp. 24–29.

Lee, F. (1962), ‘An automatic self-checking and fault-locating method’, IRE Transac-
tions on Electronic Computers EC-11(5), 649–654.

Li, W. & Shah, S. (2002), ‘Structured residual vector-based approach to sensor fault
detection and isolation’, Journal of Process Control 12, 429–443.

Lia, C.-C. & Jengb, J.-C. (2010), ‘Multiple sensor fault diagnosis for dynamic processes’,
ISA Transactions 48(4), 415–432.

Ljung, L. (1999), System Identification - Theory for the User, PTR Prentice Hall In-
formation and System Sciences Series, 2nd edn, Prentice Hall, New Jersey.

Luenberger, D. G. (1964), ‘Observing the state of linear systems’, IEEE Trans. Mil.
Electr. MIL-8, 70–80.

MacFarlane, A. & Karcanias, N. (1976), ‘Poles and zeros of linear multivariable systems:
A survey of the algebraic, geometric and complex variable theory’, International
Journal of Control 24, 33–74.

Marro, G. & Zattoni, E. (2010), ‘Unknown-state, unknown-input reconstruction

in discrete-time nonminimum-phase systems: Geometric approach’, Automatica
46, 815–822.

187
 REFERENCES

Martinez, C. J. & Wise, W. R. (2003), ‘Analysis of constructed treatment wet-

land hydraulics with the transient storage model OTIS’, Ecological Engineering
20(3), 211–222.

Massoumnia, M. A. (1986), ‘A geometric approach to the synthesis of failure detection

filters’, IEEE Transaction on Automatic Control AC-31, 839–846.

Mast, L. T., Mayper, V. & Pilnick, C. (1966), ‘Survey of Saturn/Apollo checkout

automation, spring 1965: Detailed description’, Memorandum RM-4785-NASA.

Mironovski, L. A. (1979), ‘Functional diagnosis of linear dynamic systems’, Automn

Remote Control 40, 1198–1205.

Moylan, P. (1977), ‘Stable inversion of linear systems’, IEEE Transactions on Auto-

matic Control 22(1), 74–78.

Ng, H., Chen, R. & Speyer, J. (2006), ‘A vehicle health monitoring system evaluated
experimentally on a passenger vehicle’, IEEE Transactions on Control Systems
Technology 14(5), 854–870.

Palanthandalam-Madapusi, H. & Bernstein, D. (2007), Unbiased minimum-variance

filtering for input reconstruction, in ‘Proceedings of the American Control Con-
ference (ACC’07)’, pp. New York City, USA.

Palanthandalam-Madapusi, H., Ridley, A. & Bernstein, D. (2005), Identification and

prediction of ionospheric dynamics using a Hammerstein-Wiener model with radial
basis functions, in ‘Proceedings of the American Control Conference (ACC’05)’,
Portland, OR, USA.

Papadopoulos, E. G. & Chasparis, G. C. (2002), Analysis and model-based control of

servomechanisms with friction, in ‘Proceedings of the International Conference on
Intelligent Robots and Systems (IROS 2002)’, Lausanne, Switzerland.

Park, J. & Rizzoni, G. (1994), ‘An eigenstructure assignment algorithm for the design
of fault detection filters’, IEEE Transactions on Automatic Control 39(7), 1521–
1524.

Patel, R. V. (1985), On blocking zeros in linear multivariable systems, in ‘Proceedings

of 24th Conference on Decision and Control’.

Patel, R. V. & Munro, N. (1982), Multivariable System Theory and Design, Pergamon
Press, Inc.

Patton, R. & Chen, J. (1997), ‘Observer-based fault detection and isolation: Robustness
and applications’, Control Engineering Practice 5(5), 671–682.

188
 REFERENCES

Patton, R. J. (1997), ‘Robustness in model-based fault diagnosis: The 1995 situation’,

Annual Reviews of Control 21, 103–120.

Patton, R. J. & Chen, J. (1991a), A parity space approach to robust fault detection
using eigenstructure assignment, in ‘Proceedings of European Control Conference
ECC91’, Grenoble, France.

Patton, R. J. & Chen, J. (1991b), Robust fault diagnosis using eigenstructure assign-
ment: A tutorial consideration and some new results, in ‘Proceedings of the 30th
Conference on Decision and Control’, pp. 2242–2247.

Patton, R. J. & Chen, J. (1991c), A robust parity space approach to fault diagno-
sis based on optimal eigenstructure assignment, in ‘International Conference on
Control 1991. Control ’91’, Edinburgh , UK, pp. 1056–1061.

Patton, R. J. & Chen, J. (1992), ‘Robust fault detection of jet engine sensor systems
using eigenstructure assignment’, Journal of Guidance, Control, and Dynamics
15(6), 1491–1497.

Patton, R. & Liu, G. (1994), Robust control design via eigenstructure assignment, ge-
netic algorithms and gradient-based optimisation, in ‘IEE Proceedings on Control
Theory and Applications’, Vol. 141, pp. 202–208.

Pearson, R. K. (1995), ‘Nonlinear input/output modelling’, Journal of Process Control

5(4), 197–211.

Pearson, R. K. (2003), ‘Selecting nonlinear model structures for computer control’,

Journal of Process Control 13(1), 1–26.

Pearson, R. K. & Pottmann, M. (2000), ‘Gray-box identification of block-oriented non-

linear models’, Journal of Process Control 10, 301–315.

Pierria, F., Paviglianiti, G., Caccavale, F. & Mattei, M. (2008), ‘Observer-based sensor
fault detection and isolation for chemical batch reactors’, Engineering Applications
of Artificial Intelligence 21, 1204–1216.

Pottmann, M. & Pearson, R. K. (2006), ‘Block-oriented NARMAX models with output

multiplicities’, AIChE journal 44(1), 131–140.

Putra, D. (2004), Control of Limit Cycling in Frictional Mechanical Systems, PhD

thesis, Technische Universiteit Eindhoven, Eindhoven.

Raff, T., Menold, P., Ebenbauer, C. & Allgöwer, F. (2005), A finite time functional ob-
server for linear systems, in ‘Procedings of the 44th IEEE Conference on Decision
and Control and the European Control Conference 2005’, Seville, Spain.

189
 REFERENCES

Reliability, Availability, and Maintainability Dictionary (1988), ASQC Quality Press,

Milwaukee.

Rocha-Cozatl, E., Moreno, J. A. & Vande Wouwer, A. (2012), Application of a

continuous-discrete unknown input observer to estimation in phytoplanktonic cul-
tures, in ‘Preprints of the 8th IFAC Symposium on Advanced Control of Chemical
Processes’, Singapore.

Sain, M. & Massey, J. (1969), ‘Invertibility of linear time-invariant dynamical systems’,

IEEE Transactions on Automatic Control 14(2), 141–149.

Shen, L. & Hsu, P. (1998), ‘Robust design of fault isolation observers’, Automatica
34(11), 1421–1429.

Siahi, M., Sadrnia, M. A. & Darabi, A. (2009), ‘A new method for observer design using
eigenstructure assignment and its application on fault detection and isolation’,
World Applied Sciences Journal 6(1), 100–104.

Simani, S., Fantuzzi, C. & Patton, R. J. (2002), Model-based fault diagnosis in dynamic
systems using identification techniques, Springer-Verlag, London, UK.

Söderström (2007), ‘Errors-in-variables methods in system identification’, Automatica

43(6), 939–958.

Sumislawska, M. (2009), Prediction of Limit Cycles in Hot Strip Mill Gauge Con-
trol, M.Sc. thesis, Control Theory and Applications Centre, Coventry University,
Coventry, UK.

Sumislawska, M., Burnham, K. J., Hearns, G., Larkowski, T. M. & Reeve, P. J. (2010),
Parity equation-based friction compensation applied to a rolling mill, in ‘Proced-
ings of the UKACC International Conference on Control 2010’, Coventry, UK,
pp. 1043–1048.

Sumislawska, M., Burnham, K. J. & Larkowski, T. M. (2010), Design of unknown input

estimator of a linear system based on parity equations, in ‘Procedings of the XVII
International Conference on Systems Science’, Wroclaw, Poland, pp. 81–90.

Sumislawska, M., Larkowski, T. & Burnham, K. J. (2010a), ‘Parity equations-based un-

known input estimator for multiple-input multiple-output linear systems’, Systems
Science 36(3), 49–56.

Sumislawska, M., Larkowski, T. & Burnham, K. J. (2011a), Design of unknown in-

put reconstruction filter based on parity equations for errors-in-variables case, in
‘Proceedings of the 18th IFAC World Congress’, Milan, Italy, pp. 4272–4277.

190
 REFERENCES

Sumislawska, M., Larkowski, T. & Burnham, K. J. (2012), Unknown input reconstruc-

tion observer for Hammerstein-Wiener systems in the errors-in-variables frame-
work, in ‘Proceedings of the 16th IFAC Symposium on System Identification’,
Brussels, Belgium, pp. 1377–1382.

Sumislawska, M., Larkowski, T. M. & Burnham, K. J. (2010b), Design of unknown

input reconstruction algorithm in presence of measurement noise, in ‘Procedings
of the 8th European ACD2010 Workshop on Advanced Control and Diagnosis’,
Ferrara, Italy, pp. 213–216.

Sumislawska, M., Larkowski, T. M. & Burnham, K. J. (2011b), Design of parity equa-

tions using right eigenstructure assignment, in ‘Proceedings of International Con-
ference on Systems Engineering’, Las Vegas, USA, pp. 367–370.

Sumislawska, M., Reeve, P. J., Burnham, K. J., Pozniak-Koszalka, I. & Hearns, G.

(2009), Computer control simulation and development with industrial applications,
in ‘Procedings of the 9th Polish-British Workshop’, Sienna, Poland, pp. 230–239.

Szabo, Z., Gaspar, P. & Bokor, J. (2005), Reference tracking for Wiener systems using
dynamic inversion, in ‘Proceedings of the 2005 IEEE International Symposium on,
Mediterrean Conference on Control and Automation Intelligent Control’, Limassol,
Cyprus.

Tan, C. P., Edwards, C. & Kuang, Y. C. (2006a), Robust sensor fault reconstruction
using a reduced order linear observer, in ‘9th International Conference on Control,
Automation, Robotics and Vision (ICARCV ’06)’, Singapore.

Tan, C. P., Edwards, C. & Kuang, Y. C. (2006b), Robust sensor fault reconstruc-
tion using right eigenstructure assignment, in ‘Third IEEE International Work-
shop on Electronic Design, Test and Applications (DELTA 2006)’, Kuala Lumpur,
Malaysia.

Tan, C. P. & Habib, M. K. (2004), Robust sensor fault reconstruction for an inverted
pendulum using right eigenstructure assignment, in ‘Proceedings of the 2004 IEEE
International Conference on Control Applications’, Taipei, Taiwan.

Tan, C. P. & Habib, M. K. (2006), Fault tolerance of a flexible manipulator, in ‘9th

International Conference on Control, Automation, Robotics and Vision (ICARCV
’06)’, Singapore, pp. 1–5.

Tervo, K. & Manninen, A. (2010), Analysis of model orders in human dynamics iden-
tification using linear polynomial and Hammerstein-Wiener structures, in ‘Pro-
ceedings of 2010 International Conference on Networking, Sensing and Control
(ICNSC)’, Chicago, IL, USA.

191
 REFERENCES

Wang, J., Jiang, B. & Shi, P. (2008), ‘Adaptive observer-based fault diagnosis for
satellite attitude control system’, International Journal of Innovative Computing,
Information and Control 4(8), 1921–1929.

White, J. E. & Speyer, J. L. (1986), Detection filter design: Spectral theory and algo-
rithms, in ‘Proceedings of the American Control Conference (ACC’86)’, Seattle,
WA, USA.

Wohnam, W. M. & Morse, A. (1970), ‘Decoupling and pole assignment in linear mul-
tivariable systems: A geometric approach’, SIAM Journal of Control 8(1), 1–18.

Xie, L., Zhang, J. & Wang, S. (2006), ‘Investigation of dynamic multivariate chemical
process monitoring’, Chinese Journal of Chemical Engineering 14(5), 559–568.

Ye, L., Shi, X. & Liang, J. (2011), ‘A multi-level approach for complex fault iso-
lation based on structured residuals’, Chinese Journal of Chemical Engineering
19(3), 462–472.

Yildiz, S. K., Forbes, J. F., Huang, B., Zhang, J., Wang, F., Vaculik, V. & Dudzic, M.
(2009), ‘Dynamic modelling and simulation of a hot strip finishing mill’, Applied
Mathematical Modelling 33, 3208 – 3225.

Yiua, J. C. & Wang, S. (2007), ‘Multiple ARMAX modeling scheme for forecast-
ing air conditioning system performance’, Energy Conversion and Management
48(8), 2276–2285.

Young, P. C. (2011), Recursive Estimation and Time Series Analysis: An Introduction

for the Student and Practitioner, Springer.

Young, P. C., Behzadi, M. A., Wang, C. L. & Chotai, A. (1987), ‘Direct digital and
adaptive control by input-output state variable feedback’, International Journal
of Control 46, 1861–1881.

Young, P. C. & Sumislawska, M. (2012), A control systems approach to input estimation

with hydrological applications, in ‘Proceedings of the 16th IFAC Symposium on
System Identification’, Brussels, Belgium, pp. 1043–1048.

Zhong, M., Ding, S. X., Lam, J. & Wang, H. (2003), ‘An LMI approach to design robust
fault detection filter for uncertain LTI systems’, Automatica 39(3), 543–550.

192
Appendices

193
Appendix A

Calculation of parameters xij

The matrix X is given by:

⎡ x 0 ⎤
⎢ ⎥
⎢ ⎥
1 0 0 ⋯ 0
⎢ x 0 ⎥
11

⎢ ⎥
⎢ ⎥
z1 1 0 ⋯ 0
⎢ x 0 ⎥
21

⎢ ⎥
X =⎢ ⎥
31 0 z2 1 ⋯ 0
⎢ ⋮ ⎥
(A.1)
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋮ ⋮ ⋱ ⋮
⎢ xq1 1,1 0 0 1 ⎥
⎢ 0 ⋯ zq1 −1 ⎥
⎢ ⎥
⎢ xq1 +1,1 0 0 z q1 ⎥
⎣ 0 ⋯ 0 ⎦
Eigenvalues corresponding to the linear combinations of columns of ē, λj , j = 1, ⋯, q1 ,
must fulfil the equation:
det(λj I − X) = 0 (A.2)

i.e.
⎡ x −λ ⎤
⎢ 11 ⎥
⎢ ⎥
1 0 0 ⋯ 0 0
⎢ x ⎥
j

⎢ ⎥
⎢ ⎥
z1 − λj 1 0 ⋯ 0 0
⎢ x ⎥
21

⎢ ⎥
det ⎢ ⎥
31 0 z2 − λj 1 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋮ ⋮ ⋱ ⋮ ⋮
⎢ xq1 1,1 ⎥
⎢ 0 0 0 ⋯ zq1 −1 − λj 1 ⎥
⎢ ⎥
⎢ xq1 +1,1 z q1 − λj ⎥
⎣ 0 0 0 ⋯ 0 ⎦
⎡ ⎤
⎢ x11 − λj xq1 +1,1 ⎥
⎢ ⎥
x21 x31 ⋯ xq1 1,1 (A.3)
⎢ ⎥
⎢ ⎥
⎢ ⎥
1 z1 − λj 0 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
0 1 z2 − λj ⋯ 0 0
⎢
= det ⎢ ⎥=0
⎥
⎢ ⎥
0 0 1 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋮ ⋱ ⋮ ⋮
⎢ ⎥
⎢ 0 0 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ z q1 − λj ⎥
⎣ 0 0 0 ⋯ 1 ⎦

194
 A. Calculation of parameters xij

Using a determinant expansion by minors, the following recursive expression is ob-

tained:
⎡ ⎤
⎢ ⎥
⎢ ⎥
z1 − λj 0 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
1 z2 − λj ⋯ 0 0
⎢ ⎥
⎢ ⎥
det(λj I − X) = (x11 − λj )det ⎢ ⎥
0 1 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋱ ⋮ ⋮
⎢ ⎥
⎢ 0 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ z q1 − λj ⎥
⎣ 0 0 ⋯ 1 ⎦
⎡ x xq1 +1,1 ⎤
(A.4)
⎢ ⎥
⎢ ⎥
x31 ⋯ xq1 1,1
⎢ 1z − λ ⎥
21

⎢ 2 ⎥
⎢ ⎥
⋯ 0 0
⎢ ⎥
j

⎢ ⎥
− det ⎢ ⎥
0 1 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋱ ⋮ ⋮
⎢ ⎥
⎢ 0 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ z q1 − λj ⎥
⎣ 0 0 ⋯ 1 ⎦

The determinant of a lower triangular matrix is equal to the product of its diagonal
elements, hence the first element of (A.4) is calculated via:

⎡ z −λ ⎤
⎢ 1 ⎥
⎢ ⎥
0 ⋯ 0 0
⎢ ⎥
j

⎢ ⎥
⎢ ⎥
1 z2 − λj ⋯ 0 0
⎢ ⎥
⎢ ⎥
(x11 − λj )det ⎢ ⎥
0 1 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋱ ⋮ ⋮ (A.5)
⎢ ⎥
⎢ 0 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ z q1 − λj ⎥
⎣ 0 0 ⋯ 1 ⎦
= (x11 − λj )(z1 − λj )(z2 − λj )⋯(zq1 − λj )

The second element of (A.4) is developed as:

⎡ xq1 +1,1 ⎤
⎢ ⎥
⎢ ⎥
x21 x31 ⋯ xq1 1,1
⎢ ⎥
⎢ ⎥
⎢ ⎥
1 z2 − λj ⋯ 0 0
⎢ ⎥
⎢ ⎥
− det ⎢ ⎥=
0 1 ⋯ 0 0
⎢ ⎥
(A.6)
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋱ ⋮ ⋮
⎢ ⎥
⎢ 0 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ z q1 − λj ⎥
⎣ 0 0 ⋯ 1 ⎦

195
 A. Calculation of parameters xij

⎡ z −λ ⎤
⎢ 2 ⎥
⎢ ⎥
0 ⋯ 0 0
⎢ ⎥
j

⎢ ⎥
⎢ ⎥
1 z3 − λj ⋯ 0 0
⎢ ⎥
⎢ ⎥
= −x12 det ⎢ ⎥
0 1 ⋯ 0 0
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋱ ⋮ ⋮
⎢ ⎥
⎢ 0 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ z q1 − λj ⎥
⎣ 0 0 ⋯ 1 ⎦
⎡ x xq1 +1,1 ⎤
⎢ 31 ⎥
⎢ ⎥
x41 ⋯ xq1 ,1
⎢ 1 ⎥
⎢ ⎥
⎢ ⎥
z3 − λj ⋯ 0 0
⎢ 0 ⎥
⎢ ⎥
+ det ⎢ ⎥
1 ⋯ 0 0
⎢ ⋮ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋱ ⋮ ⋮
⎢ 0 ⎥
⎢ 0 ⋯ zq1 −1 − λj 0 ⎥
⎢ ⎥
⎢ 0 z q1 − λ j ⎥
⎣ 0 ⋯ 1 ⎦

Following this recursive procedure equation (A.2) is reformulated as:

x11 (z1 − λj ) (z2 − λj ) ⋯ (zq1 − λj ) + (−1)x21 (z2 − λj ) (z3 − λj ) ⋯ (zq1 − λj ) +

(−1)2 x31 (z3 − λj ) (z4 − λj ) ⋯ (zq1 − λj ) + ⋯ + (−1)q1 xq1 ,1 (zq1 − λj ) + (A.7)
(−1)q1 +1 xq1 +1,1 = λj (z1 − λj ) (z2 − λj ) ⋯ (zq1 − λj )

The above equation needs to be fulfilled for λj , j = 1, w, ⋯, q1 + 1, i.e.

⎡ (zk − λ1 ) (zk − λ1 ) (−1)q1 −1 (zq1 − λ1 ) (−1)q1 ⎤

⎢ ⎥
q
∏k=1 − ∏qk=2
⎢ ⎥
1 1
⋯
⎢ ∏k=1 (zk − λ2 ) (zk − λ2 ) (−1)q1 −1 (zq1 q1 ⎥
− λ2 ) (−1) ⎥
⎢
q1
− ∏qk=2
⎢ ⎥
1
⋯
⎢ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋮ ⋱ ⋮ ⋮
⎢ (zk − λq1 ) − ∏qk=2 (zk − λq1 ) ⋯ (−1) (zq1 − λq1 ) (−1)q1 ⎥
⎣ ⎦
q1
∏k=1 1 q1 −1

⎡ ⎤ ⎡ λ ∏q1 (z − λ ) ⎤
(A.8)
⎢ ⎥ ⎢ 1 k=1 k ⎥
⎢ ⎥ ⎢ ⎥
x11
⎢ ⎥ ⎢ λ ∏q1 (z − λ ) ⎥
1

⎢ ⎥ ⎢ 2 k=1 k ⎥
⎢ ⎥=⎢ ⎥
x21 2
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⋮ ⋮
⎢ ⎥ ⎢ λq1 ∏ (zk − λq1 ) ⎥
⎣ ⎦ ⎣ ⎦
q 1
xq1 +1,1 k=1

Consequently, the first column of X is given by:

⎡ x ⎤
⎢ ⎥
⎢ ⎥
⎢ x ⎥
11

⎢ ⎥
⎢ ⎥ = (Ã) ˜
˜ −1 B̃
12
⎢ ⎥
(A.9)
⎢ ⎥
⎢ ⎥
⋮
⎢ xq1 +1,1 ⎥
⎣ ⎦

where an element of Ø(i) ∈ R(q1 +1)×(q1 +1) , denoted as Øj,k is:

q1
Øj,k = (−1)k−1 ∏(zl − λj ) (A.10)
l=k

196
 A. Calculation of parameters xij

˜ ∈ Rq1 +1 , denoted as B̃
whilst the j th element of the vector B̃ ˜ is:
j

q1
j ∏ (zl − λj )
˜ =λ
B̃ (A.11)
j
l=1

197
Appendix B

Proof of Lemma 5.4

(i)
For the sake of brevity, the superscript is omitted. Matrices A∗e and Aw are equal if
their appropriate columns are equal. From equation (2.103) it holds that eigenvectors
of (A − KC) are given by:

wj = fi + [ v1(i) v2(i) ⋯ vq(i) ] β̄j

(i) (i)
i
(B.1)

(i) (i)
where β̄j is a column vector of qi parameters. The parameters β̄j conform to (2.107),
which can be reformulated as:

(A − λk I) [ v1(i) v2(i) ⋯ vq(i) ] β̄k

(i) (i)
i
(B.2)
− (A − λ(i)
qi I) [ v1 ⋯ vqi ] β̄q(i) qi )ei
= (λk − λ(i)
(i) (i) (i) (i)
v2 i

for k = 1, ⋯, qi − 1. Using the notation:

β̄j = [ βj,1 βj,2 ⋯ βj,qi ]

(i) (i) (i) (i)
(B.3)

it follows that:

βk,1 Av1 + βk,2 Av2 + ⋯ + βk,qi Avqi − λk βk,1 v1 − λk βk,2 v2 − ⋯ − λk βk,qi vqi −
βqi ,1 Av1 − βqi ,2 Av2 − ⋯ − βqi ,qi Avqi + λqi +1 βqi ,1 v1 + λqi +1 βqi ,2 v2 + ⋯ + λqi +1 βqi ,qi vqi(B.4)
= (λk − λqi +1 )ei

k = 1, ⋯, qi . From (5.62a), it follows that:

Avj = zj vj + vj−1 (B.5)

198
 B. Proof of Lemma 5.4

Recalling that v0 = ei , (B.4) can be reformulated as:

βk,1 z1 v1 + βk,1 ei + βk,2 z2 v2 + βk,2 v1 + ⋯ + βk,qi zqi vqi + βk,qi vqi −1

− λk βk,1 v1 − λk βk,2 v2 − ⋯ − λk βk,qi vqi − βqi ,1 z1 v1 + βqi ,1 ei + βqi ,2 z2 v2
(B.6)
+ βqi ,2 v1 + ⋯ + βqi ,qi zqi vqi + βqi ,qi vqi −1 − λqi +1 βqi ,1 v1 − λqi +1 βqi ,2 v2
− ⋯ − λqi +1 βqi ,qi vqi = (λk − λqi +1 )ei

Then:

− (λk − βk,1 )ei + (βk,1 z1 + βk,2 − λk βk,1 )v1 + (βk,2 z2 v2 + βk,3 − λk βk,2 )v2
+ ⋯ + (βk,qi −1 zqi −1 + βk,qi − λk βk,qi −1 )vqi −1 + (βk,qi zqi − λk βk,qi )vqi =
(B.7)
− (λqi +1 − βqi ,1 )ei + (βqi ,1 z1 + βqi ,2 − λqi +1 βqi ,1 )v1 + (βqi ,2 z2 + βqi ,3 − λqi +1 βqi ,2 )v2
+ ⋯ + (βqi ,qi −1 zqi −1 + βqi ,qi − λqi +1 βqi ,qi −1 )vqi −1 + (βqi ,qi zqi − λqi +1 βk,qi )vqi

Due to the fact that ei , v1 , v2 , ⋯, vqi are linearly independent, it holds that:

λk − βk,1 = λqi +1 − βqi ,1

βk,1 z1 + βk,2 − λk βk,1 = βqi ,1 z1 + βqi ,2 − λqi +1 βqi ,1
βk,2 z2 v2 + βk,3 − λk βk,2 = βqi ,2 z2 + βqi ,3 − λqi +1 βqi ,2
(B.8)
⋮
βk,qi −1 zqi −1 + βk,qi − λk βk,qi −1 = βqi ,qi −1 zqi −1 + βqi ,qi − λqi +1 βqi ,qi −1
βk,qi zqi − λk βk,qi = βqi ,qi zqi − λqi +1 βk,qi

for k = 1, ⋯, qi . Consequently, for any j, k = 1, ⋯, qi + 1 it holds that:

λk − βk,1 = λj − βj,1
βk,1 z1 + βk,2 − λk βk,1 = βj,1 z1 + βj,2 − λj βj,1
βk,2 z2 v2 + βk,3 − λk βk,2 = βj,2 z2 + βj,3 − λj βj,2
(B.9)
⋮
βk,qi −1 zqi −1 + βk,qi − λk βk,qi −1 = βj,qi −1 zqi −1 + βj,qi − λj βj,qi −1
βk,qi zqi − λk βk,qi = βj,qi zqi − λj βk,qi

The following notation is proposed:

x11 = λj − βj,1
x21 = −βj,1 z1 − βj,2 + λj βj,1
x31 = −βj,2 z2 − βj,3 + λj βj,2
(B.10)
⋮
xqi ,1 = −βj,qi −1 zqi −1 − βj,qi + λj βj,qi −1
xqi +1,1 = −βj,qi zqi + λj βj,qi

199
 B. Proof of Lemma 5.4

Now consider the ith column of Aw , cf. (5.93):

(A − λ1 I)w1 = Aei + Av1 + Av2 + ⋯ + Avqi − λ1 ei − λ1 v1 −

(B.11)
λ1 v 2 − ⋯ − λ1 v qi

Incorporating (B.5) into (B.11) and reorganising, the following is obtained:

(A − λ1 I)w1 = Ae − (λ1 − β1,1 )ei + (β1,1 z1 + β1,2 − λ1 β1,1 )v1 +

(B.12)
(β1,2 z2 + β1,3 − λ1 β1,2 )v2 + ⋯ + (β1,qi zqi − λ1 βk,qi )vqi

Incorporating (B.10) into (B.12)

(A − λ1 I)w1 = Ae − x11 ei − x21 v1 − x31 v2 − ⋯ − xqi +1,1 vqi

(i)
(B.13)

which is equal to the ith column of A∗e .

200
Appendix C

Demonstration of Remark 6.3

From (6.16) it follows that:

r(t) = Tr̊(t) = T ΩM (t) (C.1)

where the product term T Ω can be formulated as:

T Ω = [ Cf1 Cf2 ⋯Cfr T2 ]

⎡ α(1) ⎤
⎢ τ ⎥
(1) (1)
⎢ ⎥
0 ⋯ 0 ατ −1 0 ⋯ 0 α0 0 ⋯ 0
⎢ 0 ⎥
⎢ ⎥
(2) (2) (2)
⎢ ⎥
ατ ⋯ 0 0 ατ −1 ⋯ 0 0 α0 ⋯ 0
⎢ ⋮ ⎥
⎢ ⎥
⎢ ⎥
⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮ ⋮ ⋮ ⋱ ⋮
⎢ 0 ⎥ (C.2)
⎢ ⎥
(k) (k) (k)
⎢ ⎥
0 ⋯ ατ 0 0 ⋯ ατ −1 0 0 ⋯ α0
⎢ 0 ⎥
⎣ 0 ⋯ 0 0 0 ⋯ 0 0 0 ⋯ 0 ⎦
= [ Cf1 ατ(1) Cf2 ατ(2) ⋯ Cfr ατ(r) Cf1 ατ(1) (2) (r)
−1 Cf2 ατ −1 ⋯ Cfr ατ −1

⋯ Cf1 α0
(1) (2)
Cf2 α0 ⋯ Cfr α0
(r)
]

whilst M (t) is expanded as:

M (t) = [ µ1 (t − τ ) µ2 (t − τ ) ⋯ µk (t − τ ) µ1 (t − τ + 1) µ2 (t − τ + 1) ⋯
(C.3)
µk (t − τ + 1) ⋯ µ1 (t) µ2 (t) ⋯ µk (t) ]
T

Recall equation (6.15):

τ r
r(t) = ∑ ∑ αj Cfi µi (t − j)
(i)
(C.4)
j=0 i=1

It follows from (C.1), (C.2), and (C.3) that (C.4) is equivalent to:

r(t) = T ΩM (t) (C.5)

Thus:
r̊(t) = ΩM (t) (C.6)

201