Aalborg Universitet: Kallesøe, Carsten
Aalborg Universitet: Kallesøe, Carsten
Aalborg Universitet: Kallesøe, Carsten
Publication date:
2005
Document Version
Publisher's PDF, also known as Version of record
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
? Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
? You may not further distribute the material or use it for any profit-making activity or commercial gain
? You may freely distribute the URL identifying the publication in the public portal ?
Take down policy
If you believe that this document breaches copyright please contact us at vbn@aub.aau.dk providing details, and we will remove access to
the work immediately and investigate your claim.
Ph.D. Thesis
Aalborg University
Institute of Elec. Systems
Dept. of Control Eng.
ISBN 87-90664-22-1
March 2005
ii
To my wife Susanne
and
my three sons Emil, Mikkel and Andreas.
iii
iv
Preface and Acknowledgements
This thesis is submitted as partly fulfillment of the requirements for the Doctor of Phi-
losophy at the Department of Control Engineering at the Institute of Electronic Systems,
Aalborg University, Denmark. The work has been carried out in the period January
2002 to January 2005 under the supervision of Associate Professor Henrik Rasmussen,
Associate Professor Roozbeh Izadi-Zamanabadi, and Chief Engineer Pierr Vadstrup.
The subject of the thesis is fault detection and identification in centrifugal pumps.
The thesis is mainly aimed to investigate the newest accomplishments in fault detection
and identification, in order to obtain robust and realiable detection schemes for the cen-
trifugal pump. Even though the main focus is on the application, a number of theoretical
contributions are also obtained. These are mainly in the area of robustness analysis in
event based detection schemes, and the realization of subsystems identified using Struc-
tural Analysis.
I would like to thank my supervisors Associate Professor Henrik Rasmussen, As-
sociate Professor Roozbeh Izadi-Zamanabadi, and Chief Engineer Pierr Vadstrup for
their good and constructive criticism during the whole project. A special thank should
be given to Roozbeh Izadi-Zamanabadi for many valuable and inspiring discussions
during the project. Furthermore, I would like to thank Karen Drescher for her help in
transforming the thesis into readable english. Also, I would like to thank my wife Su-
sanne for her patience and support during the whole project period, and for her help with
my english writing.
A sincere thanks goes to Dr. Vincent Cocquempot at the USTL-LAIL in Lille,
France, for letting me visit the university, and our inspiring discussions during my stay
there. I am also grateful to the staff at the Department of Control Engineering, and my
colleagues at Grundfos for providing a pleasant and inspiring working environment, and
their helpfulness throughout the whole project.
Finally, I want to acknowledge the financial support from The Danish Academy of
Technical Sciences (ATV) and Grundfos Management A/S.
v
Preface and Acknowledgements
vi
Summary
The main subject of this thesis is Fault Detection and Identification (FDI) in centrifugal
pumps. Here, it is assumed that an induction motor is driving the centrifugal pump, and
that only electrical and hydraulic quantities are measured. A state of the art analysis
of the topic has shown that signal-based approaches are the most used approaches for
FDI in centrifugal pumps. Robustness is seldom considered in these approaches. How-
ever, robustness is a very important aspect when it comes to implementation in real life
applications. Therefore, special focus is put on robustness in this thesis.
The signal-based approaches are utilizing signal processing and/or artificial intelli-
gence to obtain knowledge about the faults in the pump. To analyse robustness in these
systems, a combination of the Failure Mode and Effect Analysis (FMEA) and the Fault
Propagation Analysis (FPA) is proposed. To enable robustness analysis using the FMEA
and FPA a so-called disturbing event is introduced. Moreover, one of the manual steps
in the FPA is automated, using an algorithm developed in this thesis. The proposed anal-
ysis method is used to identify a set of signal events, which can be used for robust FDI
in the centrifugal pump. This shows the usability of the proposed method, not only for
analysis purpose, but also as a part of the design of signal-based fault detection schemes.
The most common fault in submersible pump applications is stator burnout. In the
state of the art analysis it is argued that this kind of fault is often initiated by an inter-turn
short circuit inside the stator. To understand the impact of this short circuit, a model of an
induction motor, including an inter-turn short circuit, is derived. This model is utilized
in the design of an adaptive observer, which can estimate the states of the motor, the
speed, and the inter-turn short circuit simultaneously. The observer is incorporated in a
detection scheme, by which the size of the inter-turn short circuit and the phase, affected
by the short circuit, can be found. The detection scheme is tested on an industrial test-
bench showing the capabilities of the detection scheme on a real application.
Structural Analysis (SA) is utilized in the design of residual generators for FDI in the
mechanical and hydraulic part of the centrifugal pump. The use of the SA is two folded.
Firstly, it is used to divide the centrifugal pump model into two cascade-connected sub-
parts, enabling the design of residual generators. Secondly, it is used to identify subsys-
tems, which can be used in the derivation of residual generators.
Traditionally, the results of the SA are used in the derivation of Analytical Redundant
vii
Summary
Relations (ARR). However, here a novel realization approach is proposed. With this
approach the subsystems, found using SA, are transformed into nonlinear state space
descriptions suitable for observer designs. All unknown variables, except for the states,
are eliminated in this state space description, leaving only the stability problem to be
considered in the observer design.
The proposed realization approach is used in the derivation of three residual gener-
ators for FDI in the mechanical and hydraulic parts of the pump. The obtained residual
observers are tested on an industrial test-bench, showing that the observers are robust,
with respect to changes in the operating conditions of the pump. Likewise, the tests
shows that the observers are able to detect and identify 5 different faults in the mechan-
ical and hydraulic part of the pump.
In many real life centrifugal pump applications, only slow bandwidth sensors are
available. This means that FDI schemes, based on dynamic models of the system, are not
usable. Therefore, a detection scheme, based on the steady state model of the centrifugal
pump, is proposed. This detection scheme is derived using SA to obtain ARRs. Robust-
ness, with respect to parameter variation, is incorporated in the detection scheme, with
the utilization of the set-valued approach. This algorithm is also tested on an industrial
test-bench, and is also shown to be able to detect 5 different faults in the mechanical and
hydraulic part of the centrifugal pump. Moreover, the algorithm is shown to be robust
to the operating conditions of the pump, but not to transient changes in these operating
conditions.
viii
Sammenfatning
ix
Sammenfatning
x
Contents
Nomenclature xv
1 Introduction 1
1.1 Background and Motivation . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
xi
Contents
xii
Contents
Bibliography 179
xiii
Contents
xiv
Nomenclature
Symbols
In this thesis all matrices and vectors are written with bold letters, to distinguish these
from scalar values.
xv
Nomenclature
I Identity matrix.
rs Stator resistance.
rr Rotor resistance.
lsl Leakage inductance in the stator.
lrl Leakage inductance in the rotor.
lm Mutual inductance in the induction motor.
zp Number of pole pairs in the motor.
Rs Stator resistance matrix, Rs = diag{rs , rs }.
Rr Rotor resistance matrix, Rr = diag{rr , rr }.
Ls Strator inductance matrix, Ls = diag{ 23 lm + lls , 23 lm + lls }.
Lr Rotor inductance matrix, Lr = diag{ 32 lm + llr , 23 lm + llr }.
Lm Mutual inductance matrix, Lm = diag{ 23 lm , 23 lm }.
R0r derived rotor resistance matrix, R0r = Lm L1 1
r Rr Lr Lm .
L0s 0
derived stator inductance matrix, Ls = Ls Lm Lr Lm . 1
xvi
Nomenclature
fx , mx , hx Vector field, algebraic constraints, and output maps of the dynamic system
S.
fo , ho Vector field and output maps of an output transformed system.
fz , hz Vector field and output maps of a state transformed system.
xvii
Nomenclature
Mathematical Symbols
, Positive and negative definit respectively.
>, < Larger than and smaller than respectively.
Logical expression to the left implies logical expression to the right.
Logical or operator.
Logical and operator.
x Maximum value of x.
x Minimum value of x.
R The reals.
R+ The positive reals including zero, i.e. R+ = {x R | x 0}.
Abbreviations
xviii
Chapter 1
Introduction
This thesis considers the analysis and design of algorithms for Fault Detection and Iden-
tification (FDI) in centrifugal pumps. The aim has been to investigate methods for FDI
in centrifugal pumps, with special focus on the robustness and usability of the obtained
algorithms. This means that the algorithms must be able to detect faults under changing
operating conditions, and should be robust with respect to disturbances in the system.
1
Chapter 1: Introduction
This project was initiated by a growing need, inside Grundfos, for knowledge about
the newest methods for detection of events and faults in pumps and pump systems. This
need is based on the expectation that monitoring and control systems will be commonly
used for supervision and control, of especially larger pumps, in the future. Besides that,
pumps are sometimes returned on warranty where it has been impossible to reproduce
the fault. In these cases it would be of great interest to know what the pump has been ex-
posed to before it is returned. This knowledge could be used to improve the construction
of the pumps and user manuals to avoid unnecessary returns on warranty, and thereby
unnecessary inconveniencies for the costumer.
The most common maintenance problems and faults expected in centrifugal pumps
can be divided into three main categories,
Faults which demands maintenance, such as bearing faults, and leakage due to
sealing faults.
Severe faults, which demand replacement, such as stator burnouts, and damaged
impeller.
The first item covers normal maintenance, which, to some extend, is necessary in any
application. Likewise, the second item covers replacement of wearing parts, which also
should be expected in any pump setup, when running for long time periods. The last item
covers severe damages, normally caused by unexpected faults or by lack of maintenance.
A well designed monitoring system will be able to help a user, exposed to faults,
in any of the three mentioned categories. Traditionally, the first two categories are,
in large pump applications, handled by doing scheduled maintenance on the plant. At
these scheduled maintenance procedures, a set of predetermined wearing parts are often
exchanged to avoid future breakdowns. When using a monitoring system, maintenance
can be done on demand, which will save costs for unnecessary replacement parts, and
more important, the pump only has to be stopped for maintenance when really necessary.
For the last category, a monitoring system would be able to detect and stop the pump
before a given fault causes total breakdown of the pump. In larger pumps this would
make repair possible, meaning that a replacement of the whole pump is saved.
Different sets of sensors could be used as inputs to such a monitoring system. For
centrifugal pumps the following sensors are interesting; vibration sensors, current and
voltage sensors, absolute pressure and pressure difference sensors, flow sensors, and
temperature sensors. Of these, the current and voltage sensors, and the flow and pressure
difference sensors have been considered in this project. These sensors are all reasonably
cheep and are often already mounted in a pump system. Therefore, by using only these
sensors, no additional hardware is needed for the proposed algorithms to work. There-
fore, the implementation costs for the system is reduced considerably.
2
Section 1.2: Objectives
1.2 Objectives
The aim of the Thesis is to investigate different methods for their usability in analyz-
ing and designing FDI algorithms for centrifugal pumps. In the investigation, special
emphasis is layed on the robustness and practical usability of the obtained algorithms.
In (strm et al., 2001) it is argued that methods for FDI can be divided into two
main groups, namely the model-based and signal-based approaches. Here, the signal-
based approaches are approaches, in which signal processing and/or artificial intelli-
gence are utilized to obtain knowledge about faults in a given system. The model-based
approaches are, on the other hand, utilizing system theory to obtain knowledge about the
faults. In this thesis special focus will be put on the use of the model-based approaches,
as these approaches have inherent methods for handling disturbances. Hereby, increased
robustness of the algorithms can be obtained. However, signal-based approaches have
been widely used for fault detection in centrifugal pumps and their applications. See
Chapter 2 concerning the state of the art of the area. In most of these cases robustness
has not been considered. Therefore, a method for analyzing robustness in signal-based
FDI systems, is also considered.
1.3 Contributions
The contributions of the Thesis can be divided into two groups. The first group contains
contributions to FDI in the centrifugal pumps. The second group contains theoretical
contributions, mainly on robustness analysis of signal-based fault detection schemes
and the realization of subsystems found using Structural Analysis. In this section, first
the theoretical contributions are listed, followed by the contribution to FDI in centrifugal
pumps.
The main contributions in the theoretical areas are:
A new algorithm for cutting loops in a Fault Propagation Analysis (FPA) graph is
proposed in Chapter 4. With this algorithm and a theorem also proposed in this
thesis, the FPA is automated. This means that the only manual step is to setup the
event model.
A disturbing event is introduced as a part of the FPA in Chapter 4. With this event
it is possible to analyse the robustness of signal-based fault detection algorithms.
Two theorems are formulated, aimed to analyse robustness, based on this idea.
3
Chapter 1: Introduction
An adaptive observer for estimating inter-turn short circuit faults in the stator of
an induction motor is proposed in Chapter 5. This has been published in (Kallese
et al., 2004c).
A robust FDI scheme, based on the steady state model of the pump and set-valued
algebra, is derived in Chapter 7. The obtained algorithm depends on steady state
measurements only, making it useful in cost sensitive products.
4
Section 1.4: Outline of the Thesis
Chapter 1: Introduction
The purpose of this chapter is two-folded. Firstly, the most important ideas and terms
used in the area of Fault Detection and Identification (FDI) are introduced. Secondly,
state of the art on FDI in centrifugal pumps, as well as in induction motors, is considered.
This includes contributions from the academic world and products already on the market.
This chapter introduces the mathematical model of the centrifugal pump. This includes
a model of the induction motor driving the pump, and models of the mechanical and
hydraulic parts of the pump. The presented models are lumped parameter models, which
especially are suitable for use in model-based FDI design. Special emphasize is put on
the dynamics of the hydraulic part. Here, it is shown that the energy conversion from
mechanical to hydraulic energy, is described by a purely algebraic equation. Moreover,
it is shown that the pump dynamics can be described by adding extra mass to the rotating
parts of the pump, i.e. increasing the moment of inertia of the rotating parts of pump.
The derived model is valid under two assumptions, also stated in the chapter.
In this chapter the use of Failure Mode and Effect Analysis (FMEA) and Fault Propaga-
tion Analysis (FPA) in the design of signal-based fault detection algorithms is explored.
The FMEA and the FPA are well known analysis tools, and have been proposed as an
analysis tool in the design Fault Tolerant Control, as well as in FDI algorithms. A new
algorithm for automating parts of the FPA is proposed in this chapter. Moreover, by
introducing a so-called disturbing event in the FPA, it is shown that the robustness of
signal-based FDI algorithms can be analysed.
The chapter includes an FMEA of a general centrifugal pump, meaning that the con-
ceptual faults, expected in centrifugal pumps, are identified and analyzed. The outcome
of the FMEA is a list of possible faults in centrifugal pumps. 11 of these faults are
grouped into 7 fault groups. These 7 faults found the basis for the FDI algorithms de-
signed in this thesis. Using the FPA, different sensor combinations are analysed, aimed
to find a set of signals, which can be used in a signal-based fault detection scheme. One
of these sensor configurations is proven to work on a special designed test setup.
5
Chapter 1: Introduction
6
Chapter 2
The purpose of this chapter is two-folded. Firstly, a short introduction to the most im-
portant ideas and terms used in the area of Fault Detection and Identification (FDI) is
included. Secondly, a state of the art analysis on FDI in centrifugal pump applications
is presented. The first part is included to lighten readers of the thesis not familiar with
the concept of FDI. The second part includes both a state of the art analysis of FDI in
the centrifugal pump itself, and on the induction motor drive by which centrifugal pump
is driven. Moreover, the analysis includes contributions from both the academic world
and products already on the market.
In Section 2.1, where the concept of FDI is introduced, three different approaches to
FDI are considered. First of all, distinguishing between model-based and signal-based
FDI is considered (strm et al., 2001), and the main ideas behind both methods are
described. This is followed by an introduction of the parameter adaptation approach,
and finally the concept of residual evaluation is introduced.
In Sections 2.2 and 2.3 state of the art of FDI in respectively induction motors and
centrifugal pumps is considered. A number of different faults and detection methods
have been treated in both the induction motor and in the centrifugal pump. However,
considerable more work is done in the area of FDI on induction motors compared to the
work done on centrifugal pumps. This is mainly because of the widespread use of the
motor type. The state of the art analysis is followed by some concluding remarks, which
end the chapter.
7
Chapter 2: Fault Detection and Isolation in Pump Systems
u System
y
Figure 2.1: System with inputs u, outputs y and a fault f affecting the system.
In this system x(t) Rn contains the states, u(t) Rp contains the inputs, y(t) Rd
contains the outputs, and d(t) Rl contains disturbances, which can be interpreted
as unknown or unmeasurable inputs. This system is affected by the faults f (t) Rh
affecting the system by an added term, and the parameters f Rk affecting the sys-
tem in multiplicative manner. Here the multiplicative faults are seen as changes of the
parameter values in the system. Besides the multiplicative and additive fault effects, a
fault can change the structure of a system, meaning that the system becomes a so-called
hybrid system, where the state change is caused by the given fault.
Having the above described system in mind the fault detection problem is the task
of detecting that a fault f F has occurred in a given system, where F is the set of all
possible faults in the system, i.e. it contains all faults in f and f . The solution to the
fault detection problem is based on the set of measurements y and possibly the set of
known input signals u. When a fault is detected it is possible to state that something is
wrong in the system but not what is wrong. However, sometimes it is possible to isolate
the fault, meaning that fault fi can be distinguished for the set of possible faults F in
the system. When a fault is isolated it is possible to state where and what is wrong in
8
Section 2.1: Fault Detection and Isolation
the system (Chen and Patton, 1999; Gertler, 1998). The problem of both detection and
isolation of a fault is called the Fault Detection and Isolation (FDI) problem. If it is not
only possible to isolate the fault fi in the set F, but also possible to estimate the size of
this fault fi , the fault is said to be estimated. The three levels of complexity in the fault
detection problem, described above, are summarized below.
Fault Detection: An abnormality in a system is detected, but the type and size are un-
known.
Fault Isolation: The fault fi is identified in the set of all possible faults F. Hereby the
type of the fault is known but the size remains unknown.
Different methods can be used for detecting a fault f F. The choice of method
should be based on the type of fault, which has to be detected, and which measurements
are available. Below three main groups of approaches are described.
f
Fault detection and isolation
u y Signal r Residual v
System
processing evaluation
Fault
scenario
database
Figure 2.2: Structure of a signal based fault detection and isolation system (strm
et al., 2001).
this figure it is seen that the fault detection algorithm consists of three blocks. In the
first block, signal processing, methods from signal processing theory are utilized to ex-
tract information about the health of the system from the measured signals. The output
from the signal-processing block is sent to a unit consisting of a database and some
sorts of artificial intelligence. In Fig. 2.2 this is the residual evaluation block and the
fault scenario database block. This part of the algorithm compares the output from the
9
Chapter 2: Fault Detection and Isolation in Pump Systems
signal-processing block with predefined data sets from the database, each describing the
characteristics of a given fault. From this comparison the FDI algorithm decides if the
system is affected by a fault and if so, which one it is.
The signal-processing block often consists of frequency spectrum analysis such as
FFT-algorithms, wavelets, or higher order statistic (HOS) tools. However, it could also
be a simple limit check on the measured signal. In the decision unit, i.e. the fault
scenario database and the residual evaluation block, all kinds of methods for data evalu-
ation are used. Of these clustering techniques, neural networks, and fuzzy logic should
be mentioned. All of these are sophisticated methods for data mining. However, in most
real applications simple forms of decision logic are used.
Now considering the advantages and disadvantages of the signal-based method as
the author see it. First of all, a mathematical model is not used in this approach, which is
a huge advantage, as such a model can be difficult and even in some cases impossible to
derive. However, the drawback is the need for data from the system when it is affected
by faults, as these data should be used in the development of the fault scenario database.
Moreover, it can be difficult to ensure robustness of the FDI algorithm, as in theory all
possible operation conditions should be tested, before robustness is ensured. Of course,
simulations can solve some of these problems, but then a model is needed, undermining
one of the advantages of the approach. Considering these characteristics, this approach
must be considered most suitable for systems, which are difficult or in particular cases
impossible to describe with a mathematical model.
Definition 2.1.1 This residual signal is a signal with the following characteristics,
10
Section 2.1: Fault Detection and Isolation
u System
y
model
based res.
filter
r
Residual
evaluation
Figure 2.3: Structure of a model based fault detection and isolation system.
The model-based residual filter is also called a residual generator. For linear systems
the residual generator is well established, and is well described in the literature. Here,
only two books will be mentioned, these are (Gertler, 1998) in which the parity equation
approach is treated, and (Chen and Patton, 1999) in which observer based approaches
are considered. Else, see for example (Patton and Chen, 1997; Frank and Ding, 1997)
and references included. The aim in the design of the residual generator is to be able
to create residuals with properties as defined in Definition 2.1.1, where the residuals are
not influenced by the disturbance on the system d, see (2.1). When this is possible the
residual generator is said to be robust. Different design approaches have been used to
obtain robustness. Of these should be mentioned, the unknown input observer approach
(Chen and Patton, 1999, Chap. 3), the eigenvalue assignment approach (Chen and Pat-
ton, 1999, Chap. 4), the geometric approach (Massoumnia et al., 1989), and the standard
formulation approach (Stoustrup and Niemann, 2002).
It was stated before that the residual generator for linear systems is well established.
This is not the case for non-linear systems in fact a lot of research is going on in this
field. In (Garcia and Frank, 1997) an overview is given. To mention some newer results,
for example, the geometric approach is extended to the design of residual filters for non-
linear systems in (De Persis and Isidori, 2001), and in (Stoustrup and Niemann, 1998,
2002) the internal model control approach is used to handle the non-linear parts of the
system. Moreover, the derivation of analytical redundant relations, based on structure
analysis, is described in (Blanke et al., 2003). This approach can be seen as an extension
of the parity equation approach to nonlinear systems.
11
Chapter 2: Fault Detection and Isolation in Pump Systems
r = 0 ,
where 0 are the expected parameter values and are the estimated values. The param-
eters can either be estimated by an extended Kalman filter (Del Gobbo et al., 2001) or
by means of adaptive observers (Xu and Zhang, 2004; Jiang and Staroswiecki, 2002).
With both of these approaches the states of the system and the parameters containing
fault information are estimated simultaneously. System identification, as it is presented
in (Ljung, 1999), can also be used to estimate the parameter values of the system online.
This approach is explored in for example (Isermann, 1997).
12
Section 2.2: FDI in the Induction Motors
Even though most faults fall into one of these three groups, mechanical faults such as
miss alignment and rub impact between stator and rotor are also considered. In the
following two sections detection of mechanical faults are considered first, followed by
an analysis of the used method in detection of electrical faults.
The current spectrum analysis is explored in (Eren and Devaney, 2001; Schoen et al.,
1995; Benbouzid, 2000). In (Eren and Devaney, 2001) Wavelets are used for analysing
the stator current at start up, to detect bearing faults, and in (Schoen et al., 1995) the
current spectrum, during steady state operation, is used for the same purpose. In (Ben-
bouzid, 2000) an overview is given over different signal analysis methods for current
spectrum analysis. Here such different methods as the FFT, wavelets, and Higher Order
Spectrum (HOS) analyses are considered. The obtained current spectrums are used in
the detection of bearing faults and other mechanical faults.
The spectrum analysis of vibrations is explored in (Li et al., 2000; Chow and Tan,
2000; Stack et al., 2002). In (Li et al., 2000) the signal of vibrations is transformed
into a frequency spectrum, creating attributes used as input to a neural network. The
neural network is then used to map the attributes into fault types. In (Chow and Tan,
2000; Stack et al., 2002) Higher Order Spectrum (HOS) analysis is used to extract fault
information from the signal of vibrations. Also model-based methods have been used
in the detection of mechanical faults. This is explored in (Loparo et al., 2000) where a
mathematical model of the mechanical part of the motor is developed, and used in a FDI
scheme.
In commercial products the analysis of the vibration spectrum is the mostly used
approach for detection of mechanical faults. Companies such as SKF, and Brel & Kjr
offer hand held or stationary vibration analysers, for use by the maintenance staff. Here
the spectrum of vibrations is shown, leaving it to the user to interpret the signal, and
thereby conclude if there is a fault in the motor or not. To help the maintenance staff
supervising the frequency spectrum, it is normally possible to set alarm thresholds on
parts of this spectrum.
13
Chapter 2: Fault Detection and Isolation in Pump Systems
Even though, analysis of vibration signals is considered the most used method for
detection of mechanical faults in electrical motors, advanced motor protection units can
detect mechanical faults to some extend too. Such motor protection units are offered by
for example Siemens, Rockwell Automation, and ABB. With these motor protection units
it is possible to detect faults such as blocked rotor, and high temperature, which could
be caused by mechanical faults.
14
Section 2.3: FDI in Centrifugal Pumps
lowski and Ritchie, 2000) and in (Bellini et al., 2001) the FFT of the Park transformed
current, is used to extract fault information. However, not only the FFT is used in signal-
based detection of rotor faults. For example in (Ye et al., 2003; Ye and Wu, 2001; Cu-
pertino et al., 2004) the discrete Wavelet and the Wavelet Package transforms are used
for analysing the motor current.
In industry intelligent motor protection units are commercial available. These Motor
protection units can detect ground faults and overcurrent, which again can be initiated
by inter-turn short circuits. These kinds of motor protection units are available from for
example Siemens, Rockwell Automation, and ABB. Moreover, offline analysis tools are
available from for example Baker Instrument Company, which is able to detect inter-turn
short circuits in stators of electrical machines, directly.
Bloked impeller.
Defect impeller.
Bearing faults.
15
Chapter 2: Fault Detection and Isolation in Pump Systems
16
Section 2.4: Discussion
Kenull et al., 1997). In (Perovic et al., 2001) the current spectrum is used as input to
a fuzzy Logic based classifier. This classifier is used for identification of cavitation,
clogging, and damaged impeller. In (Mller-Petersen et al., 2004; Kenull et al., 1997)
fault detection in submersible pumps is considered, based on different signal processing
philosophies.
Also the model-based approach has been used for fault detection in centrifugal
pumps (Dalton et al., 1996; Liu and Si, 1994; Wolfram et al., 2001). In (Liu and Si,
1994) a linearised model of the pump is used in the design, and the considered faults
are; motor faults and the efficiency of the pump hydraulic. In (Dalton et al., 1996)
also a linearised model of the pump is used. In this case a two-pump system is treated.
Clogging faults and faults in connection with the valves in the two pump system are
considered. In (Wolfram et al., 2001) a nonlinear model of the pump, described by a
Neuro-Fuzzy model, is used in the design. Here, faults such as various sensor faults,
leakage, clogging, cavitation, bearings faults, and impeller faults are considered.
Different protection units are commercial available. For example the monitoring
unit CU3 for protection of submersible application is offered by Grundfos. However,
this is not the only commercial available monitoring unit, as both KSB and ITT offer
monitoring units, too. KSB has just launched the PumpExpert unit, which enables detec-
tion of cavitation, bearing faults and worn impeller. Moreover, dry running protection
is included. The identification approach is based on a fault tree as described in (Koll-
mar et al., 2000a). Likewise, ITT has launched a set of monitoring units with the brand
PumpSmart. This monitoring unit is based on power level protection, and does not in-
clude adjustment to different operating points of the pump. With the adjusted alarm
levels it is only possible to detect faults with major impact on the power level. Beside
these advanced monitoring units, ABB offers a system for data collection in pump ap-
plication. With this system data measured at any given application location is made
available on a website, and evaluations of trend curves are performed.
Also special sensors for seal protection exist. Burgmann, a seal producing company,
offers a life protection system for their special designed seals, and Grundfos offers the
humidity sensor LiqTec, for protecting water lubricated seals. The company Wilo is also
working in this area, as the publication (Greitzke and Schmidthals, 2000) describes a
proposed seal protection system.
Beside the products mentioned above, most centrifugal pumps with imbedded elec-
tronic control units, do offer some kind of fault detection and protection. Likewise, for
larger pump systems customized designed monitoring systems can be available.
2.4 Discussion
In this chapter the FDI problem is introduced. Two different approached are considered,
these are the signal-based fault detection approach and the model-based fault detection
approach. It is argued that the signal-based approach has its advantages when a model
of the system is not available. However, robustness properties are difficult to establish.
17
Chapter 2: Fault Detection and Isolation in Pump Systems
For the model-based approach it is argued that the advantages are in its ability to obtain
robustness from theoretical consideration, and the drawback is the need for reasonable
good models.
This introduction is followed by a state of the art analysis of Fault Detection and
Identification (FDI) in centrifugal pump applications. Centrifugal pumps are mostly
driven by induction motors, therefore state of the art in the area of FDI in induction
motors is also considered. Here, it is stated that the most common faults in induction
motors are bearing faults. However, from internal data at Grundfos it is known that stator
burnouts are one of the main reason for faults in submersible pumps. Both signal-based
and model-based approaches have been used for detecting stator faults. The signal-
based approaches are mostly concerned in finding fault signatures in the stator current.
The model-based approaches are mostly based on steady state impedance models of the
machine. This basically means that robustness with respect to dynamic changes in the
motor speed and motor current is not considered.
From the state of the art analysis of FDI in centrifugal pumps it is seen that differ-
ent centrifugal pump faults are considered, and that different methods are used for their
detection. However, the model-based approach is fare less used than signal-based meth-
ods. This might be due to the nonlinear nature of the centrifugal pump model. It is well
known that frequency converters, making it possible to optimize the operating point of
the pump, are used more and more often as drives for centrifugal pumps. However, this
means that the detection algorithms should not only be robust with respect to changes in
the hydraulic resistance, i.e. the flow through the pump, but also to speed changes. This
has not be considered in any of the presented papers.
18
Chapter 3
In this chapter the mathematical model of the centrifugal pump is presented. The chapter
starts by describing the mechanical construction of a standard centrifugal pump. Here it
is argued that the model of the pump can be divided into three subparts,
The induction motor driving the pump.
The hydraulics of the pump.
The mechanical parts of the pump.
The induction motor is modelled using a so-called dq-model of the motor dynamics.
This type of model is extensively described in the literature. The description presented
here is based on (Krause et al., 1994; Kazmierkowski, 1994; Novotny and Lipo, 1996).
The steady state performance of the hydraulics of the centrifugal pump is extensively
described in the literature too, (Sayers, 1990; Stepanoff, 1957) and others. Here this
steady state description is extended to cover the dynamics of the centrifugal pump as
well as the steady state operation, making it particular suitable in model based FDI
algorithms. The same approach is in (Gravdahl and Egeland, 1999) used for modelling
a centrifugal compressor, but here the dynamics are neglected. Dynamics of centrifugal
pumps are treated in for example (Bka and Halsz, 2002).
In this work the model is derived using the control volume approach (Roberson and
Crowe, 1993). The derived model expresses the theoretical performance of the impeller.
To obtain a model describing the performance of a real pump extra pressure losses are
added to the theoretical model (Sayers, 1990; Stepanoff, 1957). The obtained model
describes the performance of a single impeller. However, it is shown that the same
model structure also describes the performance of a multi stage pump.
The mechanical part of the pump is modelled using simple considerations based on
Newtons second law. The frictions losses in the bearing and seals are modelled by a
simple linear friction term, as the friction losses are very small compared to the torque
necessary the drive the pump, and therefore are not important in the model.
19
Chapter 3: Model of the Centrifugal Pump
The first section of this chapter contains a description of the mechanical construction
of the centrifugal pump. The second section describes the induction motor model, and
the third section contains the derivation of the model modelling the hydraulics of the
centrifugal pump. The fourth section presents the mechanical model, and in the fifth
section each of the submodels, derived in the previous sections, are composed into the
final nonlinear state space model of the centrifugal pump. Finally, concluding remarks
end the chapter.
20
Section 3.2: Model of the Electrical Motors
Electrical motor
Shaft connection
Centrifugal pump
Impellers
Inlet Outlet
21
Chapter 3: Model of the Centrifugal Pump
iabc
Hydrualic Application
Hydraulic
application
Q
Figure 3.2: Blockdiagram depicting the connection between the different parts of the
centrifugal pump.
22
Section 3.2: Model of the Electrical Motors
rs
+ dsb
v sb i sb
dt
- zpr
drb
dt i ra
dsa
rr rs i sa
i rb dra dt
rr dt +
rr v sa
-
drc i rc
dt
-
v sc dsc
i sc
dt
+
rs
Figure 3.3: The electrical circuit diagram of the induction motor, placed in the mechan-
ical structure of the induction motor. The three outer circuits are the circuits of the three
phases of the stator and the inner circuit is the rotor circuit.
rs = r s I rr = rr I ,
where rs and rr are the electrical resistances in respectively the stator and rotor coils,
see Fig. 3.3.
The flux linkages sabc and rabc in (3.1) are given by the following expression.
Here the assumption concerning saturation, described in the start of this section, is used.
where lls and llr are the leakage inductances in the stator and rotor windings respec-
tively, and lm is the mutual inductance and is given by
cos(e ) cos(e + 2 4
3 ) cos(e + 3 )
lm (e ) = lm cos(e + 4 3 ) cos(e ) cos(e + 2 3 )
, (3.3)
2 4
cos(e + 3 ) cos(e + 3 ) cos(e )
where lm is a constant and e = zp r is the electrical angle between the stator and rotor
phases, see Fig. 3.3.
23
Chapter 3: Model of the Centrifugal Pump
where subscript s and r denote the stator variables and rotor variables respectively, and
zp r is the electrical angle between the stator and rotor circuit, see Fig. 3.3. In this
transformations Tdq0 is given by
2 4
cos() cos( + 3 ) cos( + 3 )
2
Tdq0 () = sin() sin( + 3 2
) sin( + 3 4
) .
3 1 1 1
2 2 2
Equations (3.1) and (3.2), describing the electrical and magnetic system of the in-
duction motor respectively, are transformed using the transformation (3.4). Hereby the
following description of the induction motor is obtained,
d
vsdq = Rs isdq + dtsdq
vs0 = rs is0 + ddts0
d (3.5)
0 = Rr irdq + dtrdq zp r J rdq
dr0
0 = rr ir0 + dt ,
dzp r
where zp r = dt , as r is a function of time. The flux linkages in these equations
are given by
In (3.5) and (3.6) the dq0-space is split into dq-coordinates and 0-coordinates of reasons,
which will become obvious in the following. The parameter matrices Rs , Rr , Ls Lr ,
and Lm do all have a diagonal structure, and are given by,
Rs = diag{rs , rs } Rr = diag{rr , rr }
Ls = diag{ 32 lm + lls , 32 lm + lls }
Lr = diag{ 32 lm + llr , 32 lm + llr }
3 3
Lm = diag{ 2 lm , 2 lm }
0 1
J= .
1 0
24
Section 3.2: Model of the Electrical Motors
Rewriting (3.5) and (3.6), and using the magnetizing current imdq defined such that
it fulfils the equation rdq = Lm imdq , the induction motor model becomes
disdq
L0s = (Rs + R0r ) isdq + (R0r zp r JL0m ) imdq + vsdq (3.7a)
dt
dis0
lls = rs is0 + vs0 (3.7b)
dt
di mdq
L0m =R0r isdq (R0r zp r JL0m ) imdq (3.7c)
dt
dr0
llr = rr r0 , (3.7d)
dt
where
R0r = Lm L1 1
r Rr Lr Lm L0s = Ls Lm L1
r Lm L0m = Lm L1
r Lm ,
i sc - +
v tc +
+
i sb
+
i tc i sc
dsb
+
i sa
rs dt
v sc v sb
+
v sc
rs rs rs dsc
v sb
dt rs
+ i tc + i tb + i ta v sa
dsa dsb dsc dsa i sb -
v tc v tb v ta
- dt - dt - dt - + - rs dt
v ta
- - -
+
i ta i sa
v sa -
+
v0
i 0 =0
- - +
v tb
i tb
Figure 3.4: Example of the stator windings connected in a Y- and -connection respec-
tively. The stator circuit in this figure is the same as the one shown in figure 3.3.
T
and the terminal currents itabc = ita itb itc are measurable. Therefore a relation-
ship between the voltages and currents at the terminals of the motor, and the voltages
and currents in (3.1) must be defined. This relationship is, in the following, established
for a Y- and -connected motor respectively.
25
Chapter 3: Model of the Centrifugal Pump
In the case of a Y-connected stator, the mappings between the stator variables and the
terminal variables are given by,
T
where 1 = 1 1 1 , and v0 is the voltage between the star point of the supply and
the star point of the stator circuit, see Fig. 3.4(a). Transforming these equations to
dq0-coordinates the description becomes,
Moreover it is seen from Fig. 3.4(a) that the relationsship 0 = isa + isb + isc must hold
for this circuit, therefore the currents is0 = it0 = 0.
Using the expression vs0 = vt0 v0 , obtained above, and the fact the is0 = 0, the
expression of the zero sequence quantities described by (3.7b) becomes,
0 =vt0 v0 . (3.8)
This shows that (3.7b) has no impact on the motor performance, therefore it can be
excluded from the final model of a Y-connected induction motor.
From the above argumentation the Y-connected induction motor is modelled by the
following set of equations,
disdq
L0s = (Rs + R0r ) isdq + (R0r zp r JL0m ) imdq + vtdq (3.9a)
dt
dimdq
L0m =R0r isdq (R0r zp r JL0m ) imdq , (3.9b)
dt
Remark 3.2.1 From the expression of the star point voltage (3.8) it is seen that the
start point voltage of the induction motor v0 equals the star point voltage of the supply
vt0 in the no-fault case. This can be used for fault detection in Y-connected induction
motors,
R whenever the star point voltage of the motor is measured. This scheme uses that
RT
(v s0 vt0 )2 dt 6= 0 when a fault has happend in the motor, and in the no-fault case
(v vt0 )2 dt = 0 (Tallam et al., 2002).
T s0
26
Section 3.2: Model of the Electrical Motors
In the case of a -connected stator the mappings between the stator variables and the
terminal variables are given by,
vsdq0 = Tdq0 K T1
dq0 vtdq0 itdq0 = Tdq0 KT T1
dq0 isdq0 .
This shows that vs0 = 0 despite of the value of vtdq0 , and it0 despite of the value of
isdq0 .
Using the fact vs0 = 0 in (3.7b) it becomes,
dis0
lls = rs is0 .
dt
This shows that limt is0 = 0 for every possible operating conditions. Moreover is0
is not affecting the rest of the model equations. Therefore (3.7b) can be excluded from
the final model of a -connected induction motor.
From the above argumentation the -connected induction motor is modelled by the
following set of equations,
disdq
L0s = (Rs + R0r ) isdq + (R0r zp r JL0m ) imdq + Bv vtdq (3.10a)
dt
dimdq
L0m =R0r isdq (R0r zp r JL0m ) imdq , (3.10b)
dt
27
Chapter 3: Model of the Centrifugal Pump
28
??
?
Section 3.3: The Hydraulic Part of the Centrifugal Pump
??
? ?
Volute
Outlet
??
?
Diffuser
Impeller
??
?
Diffuser
Impeller Impeller Impeller
eye eye
Figure 3.5: The mechanical construction of a single stage in a centrifugal pump. To the
left-hand side a top view is shown and to the right-hand side a side view is shown.
Qp Hp
Hydraulic part
r Tp
of the pump
Figure 3.6: Definition of the flows and pressures used in the definition of the losses in
the pump.
It is possible to obtain more accurate models of the centrifugal pump using finite
volume solutions of Navier-Stokes equations. With these solutions it is not only possible
to calculate the pressure between the inlet and outlet of the pump, but also possible to
calculate the pressure distributions inside the pump. However, these solutions include
heavy calculations and are therefore not usable in the design of FDI algorithms.
29
Chapter 3: Model of the Centrifugal Pump
|U| = rr , C is the liquid speed, and W is the liquid speed in proportion to the impeller,
all defined at the same radius r. Moreover the liquid speed C is divided into a tangential
and radial component named Cx and Cr respectively.
Cx
U
C
Cr
r
W
Eye
Figure 3.7: The impeller of a centrifugal pump. At a given radius r the liquid speed C,
the impeller speed U, and the liquid speed in proportion to the impeller W are shown.
Moreover the liquid speed C is divided into a tangential and radial component named
Cx and Cr respectively.
The torque acting on the impeller due to the liquid inside the impeller must be equal
to the change of momentum of the liquid inside the impeller. As the liquid at all time is
flowing through the impeller there must be a change of momentum from the liquid flow-
ing into the impeller to the liquid flowing out of the impeller. This change of momentum
is due to the increase in the liquid speed C2 C1 , where C1 and C2 are the liquid
speed at inlet and outlet of the impeller respectively. C1 and C2 are both functions of
the impeller angular speed r and the volume flow Qi , meaning that the mathematical
expression describing this part must be on the form
T = ft (Qi , r ) ,
where T is the torque component, Qi is the volume flow and r is the angular speed of
the impeller. The energy added to the liquid due to this torque component is converted
into static pressure in the diffuser and volute. Therefore the expression of the pressure
produced by the pump must be on the form
Hi = fh (Qi , r ) ,
30
Section 3.3: The Hydraulic Part of the Centrifugal Pump
momentum. Now recognizing that the velocity triangle in Fig. 3.7 will be fulfilled at
all time, and that the direction of the liquid speed in proportion to the impeller W is
fixed due to the impeller blades. This means that changes of the tangential speed of
the impeller U will induce changes in Cx , and therefore the momentum of the liquid
inside the impeller is changed. Moreover, Cx will also change if the radial speed of the
liquid Cr is changed. This change is explained by the velocity triangle, which has to be
fulfilled all the time. This means that a change in the volume flow Qi will change the
momentum of the liquid inside the impeller, as Cr is proportional to Qi .
From the above consideration the model describing the torque acting on the impeller
due to the liquid, must be on the form,
dQi
Ti = ft (Qi , r ) + ft0 (Qi , r , d 00
dt ) + ft (Qi , r , dt )
r
,
where Ti is the impeller torque, Qi is the volume flow, r is the angular speed of the
impeller and fh , fh0 and fh00 are functions to be decided. In fact in Section 3.3.2 the
control volume approach is used to derived an expression of the torque. Here it is shown
that the torque is described by,
dr dQi
Ti = at2 Q2i + at1 r Qi + at0 r2 + JMv KQ ,
dt dt
where at2 , at1 , at0 , JMv and KQ are constants described in Section 3.3.2. An expression
of the pressure produced by the pump is derived in Section 3.3.3. Here it is shown that
the pressure expression becomes,
where ah2 , ah1 and ah0 are constants described in Section 3.3.3.
where all Tj are external torques on the control volume, r is a radius vector, and c is the
fluid speed at the radius vector r. C is the fluid speed in proportion to a infinitesimal
control surface dA and dA is an infinitesimal area vector describing the direction and
size of the infinitesimal control surface dA. Finally is the mass density of the liquid
and dV is an infinitesimal volume. The subscribes cs and cv on the intergrals denote
the control surface and the control volume respectively. The first term on the right-hand
31
Chapter 3: Model of the Centrifugal Pump
side of (3.12) describes the momentum change due to the liquid flow and the second
term describes the dymanics.
To be able to utilize (3.12) for calculating the torque acting on the impeller, a control
volume must be defined. As only the overall performance of the impeller is of interest
????
in this work the control volume is chosen as the inside of the impeller, see Fig. 3.8(a).
The control V is represented by the hatched area with a uniform hight h and the control
????
???? ????
C2
Zoom of the control
????
volume Impeller
C Cx outlet
C1
outlet
???? ????
Inlet
R2 Cr
dm A1 Impeller U
dm inlet r
R1
V
A2 W
dm
Impeller
blad
(a) The control volume defined on the impeller. (b) The speed triangle at a given point of the im-
The hatched area defines the control volume V peller blad. c is the fluid speed, w is the fluid speed
with a uniform hight h, and dm is a small in proportion to the impeller, and u is the tangential
mass entering and leaving the impeller in time dt speed of the impeller.
through the areas A1 and A2 respectively.
Figure 3.8: Definition of the control volume and fluid speed inside the control volume.
areas A1 and A2 are the inlet and outlet surface of the impeller respectively.
At a given radius r = |r| the velocity triangle in Fig. 3.8(b) can be assumed equally
positioned proportional to the radius vector r for all angles of r. Therefore the cross
product r c is constant for all angles of r and can be expressed by,
r c = rcx , (3.13)
where r = |r| and cx is defined in Fig. 3.8(b). cx is in general a function of the radius r.
As the liquid is assumed incompressible the inlet mass flow must equal the outlet
mass flow. This means that,
Z Z
dm
C2 dA = C1 dA = , (3.14)
A2 A1 dt
where dm
dt is the mass flow, Ci is the speed of the fluid at the infinitesimal surface dA and
dA is an infinitesimal area vector describing the direction and size of the infinitesimal
surface dA.
From Fig. 3.8(a) it is seen that the infinitesimal volume dV equals,
dV = 2rhdr , (3.15)
32
Section 3.3: The Hydraulic Part of the Centrifugal Pump
where r is the radius at a given slice of the disc forming the impeller with hight h.
Using (3.13), (3.14), and (3.15) in (3.12), the torque expression is reduced to,
Z R2
dm d
Ti = (r2 cx2 r1 cx1 ) + 2h r2 cx (r)dr , (3.16)
dt dt R1
where Ti is the component of the shaft torque created by the fluid inside the impeller.
In (3.16) the first term on the right-hand side describes the time derivative of the change
of momentum between the liquid entering and leaving the impeller. This term is the
pressure or head producing part and is in the following denoted the steady state torque
term. The second term on the right-hand side describes the torque necessary to change
the speed of the impeller when the mass of the impeller itself is not taken into account.
This term is in the following denoted the transient torque term.
where the mass flow dm dt is replaced by Qi . Equation (3.18) models the steady state
torque load of the impeller.
Beside the torque described by (3.18) and extra torque term Tf = Kn r2 (Sayers,
1990) is added. This term models the hydraulic friction due to liquid between the volute
and the impeller, see Fig. 3.5. Adding this term the final steady state expression of the
torque becomes,
where,
r2 cot(2 ) r1 cot(1 )
at2 = at1 = r22 r12 at0 = Kn .
A2 A1
33
Chapter 3: Model of the Centrifugal Pump
The first term on the right-hand side of (3.20) describes the change of momentum of the
liquid inside the impeller due to changes in the impeller speed. The second term on the
right-hand side describes the tangential change of liquid speed due to flow changes. As
it is a change of speed in the tangential direction it will affect the transient component
of the shaft torque, as it is shown in (3.20).
Hp = gH ,
where H is the head and Hp is the pressure. The theoretical head He is given by He =
P/(Qi g), where g is the gravity and P = Ti r is the power. Including this in (3.18)
the following expression is obtained,
2
r2 r2 r2 r1
He = 1 r2 cot(2 ) cot(1 ) r Qi . (3.22)
g g gA2 gA1
34
Section 3.3: The Hydraulic Part of the Centrifugal Pump
The head described by (3.22) is the theoretical obtainable head of the impeller. This head
is not equal to the head between the inlet and outlet of the pump due to head losses. The
following types of head losses are expected to influence the theoretical head (Sayers,
1990),
Slip factor, here named s .
Shock losses, here named hs .
Friction losses, here named hf .
The slip factor is an empirical factor used to account for re-circulating flow inside the
impeller. The two last head losses are due to friction and shock losses inside the impeller
and at the inlet of the impeller respectively. The head losses affect the theoretical head
as shown in the following equation,
H + h s + h f = s He , (3.23)
where H is the head between the inlet and outlet of the pump. The shock and friction
losses are given by,
hs = Ks (Qi Qd )2 hf = Kf Q2i ,
where Qd is the design flow, which is a linear expression of the angular velocity e.i.
Qd = Kd r . Ks , Kf and Kd are all constants. When including these expressions in
(3.23) the model describing the head becomes,
where H is the head produced by the pump, r is the impeller speed, and Qi is the
volume flow through the impeller. The parameters in the expression are given by,
r2 r1
ah2 = Ks + Kf ah1 = s cot(2 ) cot(1 ) Ks Kd2
gA2 gA1
2
r2 r12
ah0 = 2Ks Kd s .
g g
The slip factor s in the expression above is, as explained before, an empirical scaling
factor and is in general a non-linear function of the flow and the angular velocity, (Say-
ers, 1990). However, if this factor is assumed constant the parameters ah2 , ah1 and ah0
are also constants.
3.3.4 Leakage Flow and Pressure Losses in the Inlet and Outlet
The model equations derived in the previous part of this section describe the relations
between the speed and flow of the impeller, and the pressure generated by the impeller.
35
Chapter 3: Model of the Centrifugal Pump
The impeller is not the only component affecting the hydraulic performance of the pump,
as the fluid also passes through the inlet, the diffuser, the volute and the outlet on its way
through the pump, see Fig. 3.6. To have a complete model of the hydraulic performance
?????
??
of the pump, leakage flow and pressure losses caused by these components must be
considered.
In Fig. 3.9 the leakage flow q, the inlet and outlet flow Qp , and the impeller flow Qi
are presented.
??
???
??
????
???
??
Qp
Outlet q Qi q +
aHi
-
Qp
Inlet
- + + - - +
Houtlet Hi Hinlet
Figure 3.9: Sketch of the flows inside the centrifugal pump. Qp is the main flow com-
ponent and q is the leakage flow between the pressure and suction side of the impeller.
d
KJ q = agHi Kl q 2 ,
dt
where KJ is a constant depending on the mass of the fluid involved in the leakage, a is
a scaling constant taking into account that the pressure acting on leakage flow is not the
same as the pressure delivered by the impeller, and finally Kl models pressure losses in
the loop of the leakage flow q.
The pressure losses Hinlet and Houtlet in Fig. 3.9, at the inlet and outlet of the
pump respectively, are modelled by adding an extra pressure loss term Kc in the flow
Qp . These pressure losses are called casing losses, as the pressure losses are caused by
the casing of the pump.
If each of the pressure losses are symbolized as valves, and the theoretical pressure
gHe is symbolized as a pump without losses, the diagram in Fig. 3.10 describes the
operation of the centrifugal pump.
Including the description of the lekage flow q in the expressions of the impeller head and
torque in (3.24) and (3.21) respectively, the following model of the centrifugal pump is
36
Section 3.3: The Hydraulic Part of the Centrifugal Pump
2
K cQ
q
Kf Q i 2
H
Ks(Qi -Kd r)2 2
Hi K lq
Qi
gHe
Figure 3.10: Diagram of the centrifugal pump. The valves describe the losses inside the
centrifugal pump, and the pump symbol symbolizes generated pressure without losses.
obtained,
KJ dq 2 2
dt = g(ah2 (Qp + q) + ah1 (Qp + q)r + ah0 r ) Kl q
2
where Hp is the pressure generated by the pump, Tp is the impeller torque and Qp =
Qi q is the inlet flow of the pump.
Assumption 3.3.1 For most centrifugal pumps the following assumptions hold in the no
fault case,
1. The dynamics of the leakage flow q is at least a decade faster than the main dy-
namics of the system, i.e. dq
dt 0 almost all the time.
When these assumptions hold the effect of the leakage flow on the pump performance is
neglible.
37
Chapter 3: Model of the Centrifugal Pump
When Assumption 3.3.1 holds the model (3.25) can be simplified to become,
H = H1 + H2 + + Hn (3.27)
where n is the number of pump stages. If these n pump stages are identical, meaning
that Hi = Hi1 = = Hin and q = q1 = = qn the model of a multi stage
centrifugal pump is given by the sum of each stage,
nKJ dq 2 2
dt = ng ah2 (Qp + q) + ah1 (Qp + q)r + ah0 r nKl q
2
2 2 2 0 2
Hp = ng
a h2 (Qp + q) + ah1 (Qp + q)r + ah0
r nK c p Kc Qp
Q
2 2
Tp = n at2 (Qp + q) + at1 (Qp + q)r + at0 r +
dQp dq
nJMv d
dt nKQ
r
dt + dt .
In this model nKc expresses the pressure losses due to the guidens of the flow from one
stage to the next. Therefore, an extra casing loss term Kc0 is added to model losses at the
inlet and outlet of the pump.
If Assumption 3.3.1 still holds, which is normally the case, this model can be reduced
to,
Hp = g (n(ah2 + Kc ) + Kc0 ) Q2p + nah1 r Qp + nah0 r2
dQp
Tp = nat2 Q2p + nat1 r Qp + nat0 r2 + nJMv d
dt nKQ dt .
r
From this it is seen that the structure of the model is the same for a serie connected set
of identical stages, as for a single stage centrifugal pump.
Pump curves
Normally centrifugal pumps are described by two so-called pump curves. This is
through for both multi stage and single stage pumps. The two curves depict the vol-
ume flow versus the pressure and the power of the pump respectively. Normally the
curves are only depitch for one particular speed value, which is denoted the norminal
speed. An example of these pump curves is shown in Fig. 3.11.
38
Section 3.4: The Mechanical Part of the Centrifugal Pump
6 1500
pressure [bar] 5
power [W]
4 1000
2 500
0 0
0 1 2 3 4 0 1 2 3 4
3
flow [m /h] flow [m3/h]
(a) Volume flow versus pressure difference. (b) Volume flow versus power consumption.
Figure 3.11: Pump curves describing the performance of a centrifugal pump at nominal
speed.
The impeller.
Using Newtons second law the dynamics of the mechanical parts can be described by,
d
Jm = Te Tl ,
dt
where Te is the torque produced by the motor (3.11), Tl is the load torque created by the
impeller (3.26) and mechanical friction losses respectively, and finally Jm is the moment
of inertia of the mechanical system.
The moment of inertia of the mechanical parts Jm is given by the sum of the moment
of inertia of all rotating parts of the pump, i.e. it is given by,
Jm = Jr + Js + Ji .
Here Jr is the moment of inertia of the induction motor rotor, Js is the moment of inertia
of the shaft, and Ji is the moment of inertia of the impeller.
39
Chapter 3: Model of the Centrifugal Pump
If the friction losses Tf of the bearings and shaft seals are assumed linear, i.e. Tf =
Br , the load torque is given by,
dr dQp
Tl = Br + at2 Q2p + at1 r Qp + at0 r2 + JMv KQ .
dt dt
In this equation the first term on the right-hand side models the mechanical friction
losses, and the last term models the load torque of the impeller (3.21).
Including the load torque expression in the mechanical equations the model of the
mechanical system is obtained,
dr dQp
J = Te Br at2 Q2p + at1 r Qp + at0 r2 + KQ ,
dt dt
dQ
where J = Jm + JMv . The term dtp in the above expression means that the mechanical
system depends upon the dynamics of the hydraulic application. This dependency is in
almost all applications very small, meaning that the following assumption holds for
almost all applications.
at all time. When this assumption holds it means that the application dynamics is so
slow that its effect on the pump dynamics is negliable.
dr
J = Te Br at2 Q2p + at1 r Qp + at0 r2 .
dt
x = f (x) + Gu + m(x, w)
(3.28)
y = h(x, w) .
40
Section 3.5: Final Model of the Centrifugal Pump
where isd and isq are the motor currents in a dq-frame fixed on the stator, imd and imq are
the magintizing currents described in the same frame, and finally r is the mechanical
speed of the pump.
The inputs of the system consist of the known input vector u and the unknown input
vector w, given by,
T T
u = vsd vsq w = Qp ,
where vsd and vsq are the supply voltages of the motor described in the same dq-frame
as the motor currents, and the unknown input is the volume flow Qp through the pump.
Except for the known inputs it is assumed that three variables are measured. These
are the motor currents and the pump pressure respectively. Therefore, the measurement
vector y in (3.28) is given by,
T
y = isd isq Hp .
The matrix G, the vector fields f (x) and m(x, w) and the mapping h(x, w) in (3.28)
are presented below,
R +R0 R0 L0
sL0 r isd + L0r imd zp r Lm0 imq
R +R s s s
s 0 r0 isq + zp r L0m0 imd + R0r0 imq
Ls Ls Ls
Rr0 Rr0
f (x) = i i + z i
L0m sd L0m md p r mq
Rr0 Rr0
L0m isq zp r imd L0m imq
1 3 B
J 2 z p L m (imd isq i mq isd ) J r
1
L0s 0 0
0 1 0
L0s
G=
0 0 m(x, w) =
0
0 0 0
0 0 J1 fT (Qp , r )
isd
h(x, w) = isq ,
fH (Qp , r )
where the first four rows in the model represent the electrical part of the induction motor
and the fifth row describes the mechanical part of the pump.
In this model the expression fH (Qp , r ) describes the pressure produced by the
pump, and the expression fT (Qp , r ) describes the load torque of the pump. These
expressions are derived in Section 3.3, and are given by,
fH (Qp , r ) = g ah2 Q2p + ah1 Qp r + ah0 r2
(3.29)
fT (Qp , r ) = at2 Q2p + at1 Qp r + at0 r2 ,
41
Chapter 3: Model of the Centrifugal Pump
where ah2 , ah1 , ah0 and at2 , at1 , at0 are constant parameters found from the physical
properties of the pump, see Section 3.3.
The model presented above is not valid in the whole state space since the functions
fH (Qp , r ) and fT (Qp , r ) are only valid for positive flow and speed. Therefore, the
state space of interest for this system is given by D = {(x, w)|(x1 , , x4 ) R4 , x5
R+ , w R+ }, where x1 , , x5 are states in the state vector and w is the unknown
input signal.
3.6 Discussion
In this chapter a model of the centrifugal pump including an induction motor drive is
derived. The obtained model is a fifth-order lumped parameter system, making Fault
Detection and Identification designs, based on this model, possible. However, the model
is very nonlinear, therefore a linearized version of the model is only expected to work
in a small neighborhood around the point of linearization. The model is composed of
three sub-models describing the electrical part of the motor, the mechanical parts of the
system and the hydraulic parts of the system respectively.
In the derivation of the model the dynamics is taking into account. It is shown that
under two assumptions given in the chapter, and if the mass of the liquid inside the
impeller is added to the mass of the impeller itself, the model of the hydraulics becomes
purely algebraic. This means that the liquid inside the impeller affects the moment of
inertial in the mechanical description of the pump, and that no additional dynamics are
added to the model due to the hydraulics of the pump.
The obtained model or sub-models will in the following be used in the derivation
of FDI-algorithms. Moreover, the understanding of the system obtained through the
model, is used in the design of a test setup, where a number of hydraulic, mechanical
and electrical faults can be simulated.
42
Chapter 4
43
Chapter 4: System Analysis and Fault Modelling
are developed.
By including the notation of disturbing events in the FPA, the use of the approach
is extended to handle robustness of signal-based fault detection schemes. However, it
does not solve the problem with the manual steps in the analysis. One of these manual
steps is the cutting of loops inherent in the system under investagation (Bgh, 1997).
In this chapter an automated way of handling these loops is proposed, meaning that the
remaining manual step is to setup the event model.
The chapter starts by presenting some preliminaries on the FMEA and FPA and their
use in connection with fault detection, isolation and accommodation in Section 4.1. The
remaining parts of the chapter are devoted to the presentation of the contributions on the
FMEA and FPA obtained in this work. First, the theoretical contributions are presented,
followed by the FMEA and FPA results on the centrifugal pump. In Section 4.2 the
theoretical results are presented. They are obtained using FMEA and FPA for robustness
analysis of signal-based fault detection schemes. In Section 4.3 the results obtained by
using the ideas, developed in Section 4.2, on the centrifugal pump are presented. This
includes a list of faults expected in the system, a FMEA and FPA model of the system,
and an analysis of the signal based detection possibilities. Finally, the chapter ends
with a presentation of test results obtained on a test setup particular developed for this
propose, and some concluding remarks. This is in Section 4.4 and 4.5 respectively.
Fault modeling - step 1: In this step a qualitative model of the faults in each componet
of the system is made. This is done by dividing the system into suitable compo-
nents. In each of these components the faults and their effects are identified and
described. This is done using the Failure Mode and Effect Analysis (FMEA).
Fault propagation analysis (FPA)- step 2: In this step the propagation of the identi-
fied faults through the component of the systems is analysed. The result of this
analysis is a description of the connection between the faults in the system and a
set of interesting effects. This analysis is called Fault Propagation Analysis (FPA),
and is performed by first identifing the functional connection between the com-
ponents of the system, ending with a functional diagram (Bgh, 1997). After that
this diagram is used for the fault propagation analysis.
Detector design - step 6: In this step detectors for detecting the fault in the system are
developed. Different approaches have be used in this step. The state of the art
on the area of fault detection and isolation in pumps is presented in Chapter 2.
44
Section 4.1: Method for Fault Analysis
In this chapter the two first steps are performed on the centrifugal pump, while the rest
of the thesis is devoted to the last step.
From the above list it is seen that FMEA and FPA are used in steps 1 and 2. These
analysis tools form a systematic way of creating a high level qualitative model of the
system under consideration. The model needs only to be descriptive for the behaviour
under faulty conditions, which allow the model to be simple compared to dynamical
models.
Traditionally the FMEA and FPA have been used for analysing the fault behaviour
of a system, and to identify reconfiguration possiblities. Hereby the conditions for fault
robust control of the given system are obtained (Blanke et al., 2003). However, in this
work the purpose of introducing the FMEA and FPA is different. Here these analysis
tools are used as a first step in developing FDI algorithms. The purpose of using FMEA
and FPA is two folded, as it is used for both identifing the set of most important faults
in the system, and analysing different combinations of sensors for detection possiblities.
These sensor combinations are in this thesis denoted sensor configurations. The first of
these purposes is the normal outcome of a FMEA analysis when used in the design of
fault robust control algorithms (Blanke et al., 2003) and qualitative analysis of products.
45
Chapter 4: System Analysis and Fault Modelling
Table 4.1: An example of a typical FMEA worksheet (Blanke et al., 2003, Chap. 4). In
this example the result of analysing a pressure sensor is shown.
Item ident. Failure mode Failure cause Failure effect Risk assess-
ment
Pressure sen- Clogging Dirt Zero output High
sor
Broken sup- Mechanical Undefined Low
ply wire vibration output
Definition 4.1.1 (Fault propagation matrix) (Blanke et al., 2003, p. 78) For a given
boolean mapping M,
M : F E {0, 1}
of the finite set of component faults F onto the finite set of effects E. The fault propaga-
tion matrix is defined as follows
1 if fcj = 1 eci = 1
mi,j =
0 otherwise ,
46
Section 4.1: Method for Fault Analysis
f3
e3= e end
Comp. 3
f1 f2
e1 e2
Comp. 1 Comp. 2
Figure 4.1: The propagation of failures in a system. The failures f1 and f2 are propor-
gated through component 3, thereby their end effects are identified in e3 = eend .
The result of the FPA is a connection between faults in the system and a decided
set of end-effects. The understanding of the propagation can be used to identify where
in the system faults can be stopped in order to prevent total failure of the system. This
knowledge can then be used in the development of reconfiguration logic for fault ac-
commodation.
In Fig. 4.1 it is emphasized that the connections between the faults in the system
and the end-effects are given by simple propagation through the components of the dia-
gram. Unfortunately this is not always the case, as loops can occure in the FPA diagram
(Blanke et al., 2003, Chap. 4). An example of a FPA diagram with a loop is shown in
Fig. 4.2. Such loops arise due to the physical structure of the system, and can therefore
f3
e3= eend
Comp. 3
f1 f2
e1 e2
Comp. 1 Comp. 2
Figure 4.2: Loop example in a fault propagation diagram of a system contaning three
components.
not be avoided in the model. Instead the loops are treated by cutting the connection
somewhere in the loop, and then extend the set of faults with the cutted effects. After-
wards each of the cutted effects are analysed to decide if they could be removed from
the FPA or should be treated as an extra fault.
In the above text it is mentioned that the FMEA and FPA traditionally are used for
designing Fault Tolerant Control systems. But if the end-effects are chosen as a subset
47
Chapter 4: System Analysis and Fault Modelling
of the measurable effects in the system, the FPA can be used in the development of
FDI algorithms. An example of this is shown in (Thomsen, 2000). Here the FPA is
used for analysing sensor configurations, revealing the connection between the faults in
the system and the set of measurable signals. Hereby the usability of different sensor
configuration can be analysed. In the following this approach is further developed to
handle robustness with respect to events in the system, which should not be considered
as faults. Finally, the developed approach is used in the analysis of the centrifugal pump.
fi
ej ei
Comp. i
Figure 4.3: A single component in a FMEA. fi is the possible failures of the component,
ej is effects affecting the component and ei is the effects on the component performance
due to fi and ej respectively.
in the component, ej contains the effects affecting the component, and ei contains the
effects on the component due to fi and ej respectively. The component shown in Fig. 4.3
implies that a component in the FPA diagram can be described as defined in definition
4.2.1.
Definition 4.2.1 (Component Sdescription) The ith component ci in the system S con-
n
taning n components, i.e. S = i=1 ci , is described by,
i1
X n
X
ei Aifi fi + Aij ej + Aij ej , (4.1)
j=1 j=i+1
where
48
Section 4.2: Automated FPA
Remark 4.2.1 In Definition 4.2.1 Aij = 0 means that there does not exist a connection
from component j to component i in the FPA diagram. And Aifi = 0 means that no
faults affect the ith component.
Remark 4.2.2 This description is similar to the one used in (Jrgensen, 1995; Blanke
et al., 2003) except for opertator and matrix notation. In (Blanke et al., 2003, Chap.
4) all the matrices Aifi and Aji are lumped into one matrix Mfi . An example of such a
matrix is shown below,
fi fi
e1 e1
ei Mfi e2 = Mfi,f Mfi,e1 Mfi,e2 . . . Mfi,en e2 ,
.. ..
. .
en en
49
Chapter 4: System Analysis and Fault Modelling
The bi-partite graph Gf is a graph with m vertices U and n vertices Y. Where the vertex
ui U is associated with the fault vector fi and the vertex yj Y is associated with the
effect vector ej . The edges of Gf are defined by the following rule,
An edge exists between ui and yj if the matrix Aifj 6= 0, where the matrix Aifj is
defined in Definition 4.2.1.
A example of the graphs defined in Definition 4.2.2 are shown in Figs. 4.4 and 4.5.
The above definition defines a general graph representation of the FPA diagram. The
adjacency matrix De associeated with Ge includes the structure of the FPA diagram,
meaning that loops in the FPA diagram are seen in this matrix. Using this graph repre-
sentation it is possible to define the properties for the system to be on calculable form.
This is formulated in Defintion 4.2.3. Here the calculability is defined from the structure
of the adjacency matrix De of Ge .
where di,j {0, 1}. Moreover the vertex representing the end-effect eend must be the
last vertex in Y, i.e. the vertex yn is associated with eend .
All FPA diagrams with a graph representation as defined in Defintion 4.2.2, which do
not contain loops, can be transformed to this form by changing the order of the vertices
in Y (Shih, 1999).
If the graph representation is on the form defined in Definition 4.2.3 the following
theorem can be used for calculating the connection between faults and end-effects.
Theorem 4.2.1 Let the graph representation of a system be on the form defined in Def-
inition 4.2.3, with n vertrices in Y and m vertices in U. Then by association of the edge
T T
di,j in De with Aij , the connection between the faults f = f1T f2T fm and
the end-effect vector en is given by,
en T 1 T2 ... Tn1 I Af f (4.2)
50
Section 4.2: Automated FPA
where,
Tn1 = Ann1
k1
X
Tnk = Annk + Tni Ani
nk
i=1
and
1
Af1 A1f2 A1fm
A2f A2f2 A2fm
1
Af = . .. .. .. .
.. . . .
Anf1 Anf2 Anfm
In Af the submatrix Aifj 6= 0 if there is an edge from the vertex uj U to the vertex
yi Y else Aifj = 0.
Proof: If the graph Ge of a system, defined as in Definition 4.2.2, is on the form defined in
Definition 4.2.3 the logical structure of the system is given by,
e1 = A1f f
e2 = A21 e1 + A2f f
e3 = A31 e1 + A32 e2 + A3f f
.. (4.3)
.
n1
en1 = An n n
1 e1 + A2 e2 + + An2 en2 + Af f
en = A1 e1 + A2 e2 + + An2 en2 + An
n n n n
n1 en1 + Af f
where each Aji corresponds to dj,i in the adjacency matrix De of the digraph Ge , and f =
T
T T
f1 f2T fm , where m is the number of components affected by faults.
We search for a solution of en on the form,
en = T1 T2 Tn1 Tn Af f (4.4)
en = Tn An n
f f = I Af f
51
Chapter 4: System Analysis and Fault Modelling
meaning that,
Tn = I (4.5)
Tn1 is found by setting Aif = 0 for all i 6= n 1. From (4.3) and (4.4) it is seen that all ej
for j < n 1 are equal to zeros in this case, and en1 = An1
f f . This implies that,
en = Tn1 An1
f f = An
n1 en1
meaning that,
Tn1 = An
n1 (4.6)
where Tni propagates the effect vector eni to en . From this equations it is seen that,
n(k1)
Tnk = An
nk + Tn(k1) Ank + + Tn1 An1
nk
k1
X
Tnk = An
nk + Tni Ani
nk (4.7)
i=1
The above theorem represents a simple solution for finding the connection between fault
and end-effects in the system. The theorem uses Definition 4.2.3, which is the same
as assuming that no loops exist in the FPA diagram. Therefore, loops must be cutted
before the theorem can be used. This is a well known problem and is described in
(Blanke, 1996; Blanke et al., 2003; Bgh, 1997). In these references it is suggested that
a solution to the loop problem is to cut loops at places, which are sensible in a physical
sense.
However, it can be argued that optimal cuts would be the set of cuts, which maxi-
mizes the number of faults seen in the chosen end-effects. By intuition these cuts would
be placed such that the backward walk from the end-effect to the cut is as long as pos-
sible. As an example see Fig. 4.4, where y4 is chosen as the vertex associated with the
end-effect vector.
In this figure the edges d2 , d3 and d4 form a loop. This loop can be cutted at each
of these edges. But by cutting d4 all faults associated with the vertises y2 and y3 can
be seen in the effects associated with the vertex y4 . This is not the case if the loop is
cutted at either d2 or d3 . By doing this either all effects associated with y2 and y3 , or
the effects associated with y2 can not be seen in the effects of y4 . When d4 is cutted it
is possible to do the backward walk y4 y3 y2 , whereas in the other two cases the
52
Section 4.2: Automated FPA
y1
d1
y2 d2 y3 d3 y4
d4
Figure 4.4: An example of the graph Ge from Definition 4.2.2. This graph represents
the structure of the FPA diagram.
Algorithm 4.2.1 (The cutting algorithm) Assuming that a system is described by two
graphs, as defined in Definition 4.2.2, with n vertices yi Y and m vertices ui U,
where Y and U are ordered sets. Let Df : U Y be the adjacency matrix associated
with the graph Gf and De : Y Y be the adjacency matrix associated with the graph
Ge . Then the algorithm is as follows:
Initialization: Transform Df and De such that the vertex in Y associated with the end-
effect eend is the last vertex Y, and for all vertices at position 1 to n 1 in Y;
remove all vertices with zero colomns in De recusively, and set i = n0 such that
the vertex yi is associated with eend , where n0 is the number of vertices after
removing zero colomns.
Step 1: If there are non-zero elements above the diagonal element in the ith colomn of
De then, set these equal to zero and add the vertex yi to the set of vertices U and
add a colomn in Df corresponding to this new vertex yi U. This new colomn
must have zero elements at position i to n0 and a structure corresponding to the
ith colomn of De at position 1 to i 1.
53
Chapter 4: System Analysis and Fault Modelling
Step 2: Sort out the vertices at possition 1 to i 1 in Y such that the vertex at position
i 1 is the vertex with edges incident to one or more of the vertices at position i
to n0 and with fewest edges incident to the vetices at position 1 to i 1.
Step 3: Set i := i 1 and go to step 1.
Using this algorithm on a given graph representation, the graphs Gf and Ge are forced
to have a structure as defined in Definition 4.2.3. Therefore Theorem 4.2.1 states the
connection between fault and end-effects. The obtained logical expression is on the
form,
eend A f . (4.8)
In this expression f contains both the faults in the system, and the effects cutted using
the cutting Algorithm 4.2.1. As it is argued in (Blanke et al., 2003; Bgh, 1997), each
of the cutted effects should be analysed to check if they can be omitted in the analysis,
or should be treated as an additional fault.
Remark 4.2.3 It would be possible to define the two graphs in Definition 4.2.3 such
that each vertex in U is associated with a single fault fi and not a fault vector fi , and
likewise each vertex Y is associated with a single effect ei and not a vector of effects
ei . This approach is not chosen here as the physical meaning is somewhat lost by doing
that.
In the following example the result of using this algorithm on a small system containing
only four components is shown.
Example
Using Definition 4.2.1 a system containing 4 components is shown in Table 4.2.1. Here
it is seen that each component ci is described by the propagation matrices Aifi and Aij .
The Aij = 0 is omitted in the system representation.
54
Section 4.2: Automated FPA
The graph representation of the system in Table 4.2.1 is, according to Definition
4.2.2, given by the following two graphs,
11,f 1 0 0 0 0 0 11,4
0 0 0 0 0 12,3 0
Df =
0
De =
13,1
13,f 3 0 13,2 0 0
0 0 14,f 4 14,1 0 14,3 0
where the set of vertices U = {uf1 , uf3 , uf4 } and Y = {ye1 , ye2 , ye3 , ye4 }. In these
adjacency matrices the symbol 1j,i corresponds to a 1 in the matrix, and the subscript
i, j denotes the position of vertices joint by the edge before the transformation. Here the
edge is incident from the j th vertex and incident to ith vertex.
In these sets the vertex ufi is assosiated with the fault vector fi and likewise the
vertex yej is assosiated with the effect vector ej . The graphs are shown in Fig. 4.5.
ye1
d1,f1 ye1 d4,1 d3,1
uf1 ye4 d1,4 d
ye2 4,3 ye3
d3,f3
uf3 d2,3
ye3
d4,f4
uf4 d3,2
ye4 ye2
(a) (b)
Figure 4.5: The graphs Gf (Fig. (a)) and Ge (Fig. (b)) before edges are cutted.
Now choose e3 , e.i. ye3 Y, as the end-effect in the analysis. With this end-
effect the adjacency matrix De of the graph representation is not on the form defined in
Definition 4.2.3. Therefore the cutting Algorithm 4.2.1 is used to cut and sort edges in
Ge to obtain the form defined in Definition 4.2.3. The first run through Algorithm 4.2.1
is shown below.
with U = {uf1 , uf3 , uf4 } and Y = {ye1 , ye2 , ye4 , ye3 }. Set i = n = 4 meaning
that the last vertex in Y, ye3 , is treated in the algorithm.
55
Chapter 4: System Analysis and Fault Modelling
Step 1: Cut the edges above the ith diagonal element, add the ith vertex of Y to U, and
the cutted edges to Gf , e.i. Df . This results in,
11,f 1 0 0 0 0 0 11,4 0
0 0 0 1 0 0 0 0
Df = 0
2,3 D =
0 14,f 4 14,3 e 14,1 0 0 0
0 13,f 3 0 0 13,1 13,2 0 0
with U = {uf1 , uf3 , uf4 , ue3 } and Y = {ye1 , ye2 , ye4 , ye3 }.
Step 2: Sorting out De such that the vertex at position i 1 is a vertex with edges inci-
dent to the ith vertex, and as few as possible no zero elements above the diagonal
in De . This results in,
1f 1,1 0 0 0 0 11,4 0 0
0 0 1f 4,4 13,4 0
Df = De = 14,1 0 0
0 0 0 13,2 0 0 0 0
0 1f 3,3 0 0 13,1 0 13,2 0
with U = {uf1 , uf3 , uf4 , ue3 } and Y = {ye1 , ye4 , ye2 , ye3 }.
Step 3: Set i := i 1 and return to step 1. This means that the vertex ye2 is treated in
the next run through the algorithm.
After 5 cycils of the algorithm the following two adjact matrices are obtained,
0 14,f 4 0 14,3 14,1 0 0 0 0
11,f 1 0 0 0 0 11,4 0 0 0
Df = 0
D =
0 0 12,3 0 e 0 0 0 0
0 0 13,f 3 0 0 0 13,1 13,2 0
where U = {uf1 , uf3 , uf4 , ue3 , ue1 } and Y = {ye4 , ye1 , ye2 , ye3 }. The resulting graphs
are depicted in Fig. 4.6.
These are on the form defined in Definition 4.2.3, hence Theorem 4.2.1 can be used
to obtain relations between the faults and the end-effects. In this example the following
connection between the extended fault vector and the end-effects is obtained,
f1
f3
e3 A31 A1f1 A3f3 A31 A14 A4f4 (A32 A23 + A31 A14 A43 ) A31 A14 A41
f4 .
e3
e1
The remaining task is to analyse each cut to validate the results against the physical
proporties of the system.
56
Section 4.2: Automated FPA
uf1 d1,f1
ye1 ye1
uf3 d3,f3
d3,1
ye3
d2,3 ye4 d1,4
ye3 ye3
d4,f4 d4,3 ye2
uf4
d4,1 ye4
ye1 ye2 d3,2
(a) (b)
Figure 4.6: The graphs Gf (Fig (a)) and Ge (Fig. (b)) after edges are cutted.
where f is the set of faults in the system, d is the set of disturbing events, and Af ,
Ad are logical matrices defined as in Definition 4.1.1. This means that (4.9) is only a
factorization of (4.8). The factorization into f and d of the fault vector is formilized in
the following definition.
57
Chapter 4: System Analysis and Fault Modelling
ef j 6= edi di , dj Fd (4.10)
Definition 4.2.5 (Logical Robust Fault Detectability) The fault vector fk Ff is log-
ical detectable if the effect vector eend in the case of the fault vector fk is different from
the effect vector eend in the no fault case. This must be true for all possible combinations
of disturbing events in both the fault and in the no fault case.
From (4.10) it is seen that for a fault to be logical detectable in a robust manner,
it is necessary that at least one of the effects of the fault cannot be corrupted by any
disturbing events. The following theorem states the conditions for this to be possible.
Theorem 4.2.2 (Logical Robust Fault Detectability) Let If be the set of vectors with
only 1 element different from zero. Let fk If be a vector with only the k th element
different from zero, and let Af and Ad be defined as in (4.9). Let (Af )j,k be the j, k th
element in Af and likewise for (Ad )j,k , then the fault described by fk is logically de-
tectable iff,
nd
X
j {1, 2, , ne } : (Af )j,k = 1 and (Ad )j,i = 0 (4.11)
i=1
where ne is the number of end-effects and nd is the number of disturbing events. This
is the same as saying that fk can be distinguished from all possible combinations of
disturbing events.
th
Proof: (AP
f )j,k = 1 implies that the j element of efk Af fk is equal to 1 for fk Ff .
Moreover, n d
i=1 (A )
d j,i = 0 implies that the j th element in ed Ad d is equal to zero for
58
Section 4.2: Automated FPA
all d Fd , i.e. the j th element of e will always be equal to zero in the no fault case. This
again implies that, only faults can affect the j th element in e. Therefore, if the j th element of
e Af f + Ad d is different from zero it must be due to f . fk will cause the j th element in e to
be different from zero, therefore (4.11) implies that fk is robust detectable.
Now assume that efk 6= ed , where
efk Af fk + Ad d (4.12a)
ed Ad d . (4.12b)
This implies that there exists an element j in efk and ed , such that the j th element in efk is
different from the j th element in ed . For this to be true the j th element in ed must be equal to
zero, as the term Ad d is part of both
P (4.12a) and (4.12b). For ed to be zero for all d Fd the
j th row of Ad must equal zero, i.e. n d
i=1 (A d )j,i = 0. Therefore, for the j
th
element of efk to
th th
be different from the j element of ed , the j element of efk must be different from zero. This
implies that (Af )j,k = 1 for the fault fk to be detectable.
Remark 4.2.4 From Theorem 4.2.2 it is seen that if Fd = , meaning that no disturbing
events exist in the system, then the demand for detectability of fk reduces to (Af )j,k = 1
for some j {1, 2, , ne }.
If Theorem 4.2.2 is fulfilled for a given fault in a system, then this fault is detectable
despite of the disturbing events affecting the system.
In some cases not all disturbing events are independent. As an example there could
exist two disturbing events d1 and d2 ; d1 saying that an input to a given component is
increasing and d2 saying that the same input is decreasing. In this case d1 and d2 are
mutually excluded or d1 = 1 d2 = 0 and d2 = 1 d1 = 0. When this is the case
Theorem 4.2.2 is too restrictive.
In general such dependencies between disturbing events can be described by,
where there are disturbing events di1 to di , which exclude disturbing events dj1
to dj . The following Corollary relaxes the demands for logical robust fault detection,
when h dependencies on the form (4.13) are assumed.
Corollary 4.2.1 (Logical Robust Fault Detectability) For the cth dependency expres-
sion on the form (4.13), define + 1 fault propagation matrices, where matrices are
formed by setting the ith l column equal 0, l {1, 2, , }, and one matrix is formed
by setting the j1th , j2th , , jth columns equal 0. These matrices form the set,
59
Chapter 4: System Analysis and Fault Modelling
Iff Theorem 4.2.2 holds for a fault fk If for all Ah Ah , then the fault is logical
detectable.
Proof: Before stating the proof a set of disturbing event vectors is defined. From Theorem 4.2.2
it is obvious that all end-effects, which can be affected by disturbing events, will be affected if all
elements in d is equal to 1. Let this worst case vector be given by 1d . When there are mutual
excluded disturbing events, as given in (4.13), the vector 1d becomes too restictive. In this case
the vectors with the maximal possible number of elements equal to one, must form the set of worst
case vectors. Let this set be given by,
Dc = {d1 , d2 , , dm }
Examining (4.13) the first m 1 vectors must be formed by setting one of the elements i1 to i
equal to zero. The mth vector is formed by setting the elements j1 , , j equal to zero. From
this it is deduced that there exist m = + 1 independent vectors in Dc describing the possible
worst case disturbing event vectors, when the mutual exclusion (4.13) exists. If there exist h of
these mutual exclusions, forming h sets of vectors D1 , D2 , , Dh all possible worst case vectors
can be defined as,
D = {d | d = d1 d2 dh , where d1 D1 , d2 D2 , , dh Dh }
where is the logical "and" operator. Each of the elements in D is formed by c + 1 vectors,
meaning that the total number of vectors becomes hc=1 (c + 1).
To prove Corollary 4.2.1 recognise that Theorem 4.2.2 holds for all Ah Ah . This implies
that fk is logical robust detectable with respect to each Ah V Ah , which again implies that
ef 6= ed where ef Af fk + Ah d and ed Ah d. Ah = hc=1 Ac , where Ac Ac is on
the form described in the corollary. From the definition of Ac it is seen that for each Ac Ac
there corresponds exactly one dc Dc such that Ac 1d = Ad dc . Using this the following is true
for each Ah Ah ,
h ! h ! h ! h !
^ ^ ^ ^
Ah 1d = Ac 1d = Ac 1d = Ad dc = Ad dc = Ad d
c=1 c=1 c=1 c=1
where d D. Due to the one to one correspondency between dc Dc and Ac Ac the above
equation implies that if Theorem 4.2.2 hold for each Ah Ah , then fk is logical robust detectable
for every disturbing event vector d D. In the start of the proof it was argued that all worst case
disturbing event vectors are contained in D. Therefore, fk is detectable for all possible disturbing
event vectors.
To show sufficiency just reverse the proof.
Theorem 4.2.2 and alternatively Corollay 4.2.1 states the demands for a fault to be
robust detectable in a logical sense. If this is the case for a set of faults in a system, it
60
Section 4.3: Pump Applications
is interesting to know if the faults in this set can be distinguished from each other i.e.
efk 6= efi whenever fk 6= fi where fk , fi If . The following theorem states the
demands for this to be possible,
Theorem 4.2.3 (Logical Fault Identification) Let If be the set of vectors with only 1
element different from zero. Let fk If be the vector describing the k th fault and let
(Af )k denote the k th column in Af then the fault described by fk can be distinguished
from all other f If in a logical sense iff,
(Af )k 6= (Af )i i 6= k, i {1, 2, , n} . (4.15)
Proof: Two faults fk and fi (fk 6= fi , fk , fi If ) are distinguishable if efk 6= efi , where
efk Af fk and efi Af fi . Let fk be a vector with only the kth element different from zero.
Then efk equals the kth column in Af , i.e. (Af )k . Likewise, let fi be a vector with only the ith
element different from zero. Then efi equals the ith column in Af , i.e. (Af )i . From this it is
immediately seen that (4.15) implies that efk 6= efi whenever fk 6= fi , where fk , fi If . This
completes the proof.
Remark 4.2.5 It should be noted that the effects used for fault identification must not
be corrupted by disturbing events. Therefore
Pnd if there are disturbing events in the system
only the end-effects ej associated with i=1 (Ad )j,i = 0 should be used in Theorem
4.2.3.
Remark 4.2.6 In (Blanke et al., 2003) a methods for defining the logical connection
from the effects to the faults is given, i.e. f B e where is a special operator
defined in (Blanke et al., 2003). Using this expression it is possible to identify a given
fault from the measurable effects, whenever Theorem 4.2.3 is fulfilled for the system.
If both Theorems 4.2.2 and 4.2.3 are fulfilled for a set of faults in a system, then this set
of faults are said to be robust identifiable in a logical sense. Whenever this is the case it
is possible to measure a set of effect in the system, and from these measurements detect
and isolate the faults.
Unfortunately for many systems it is not possible to find a set of measurable effects
where both Theorems 4.2.2 and 4.2.3 are fulfilled. However, in many cases Theorem
4.2.3 is fulfilled but not Theorem 4.2.2. In these cases the problem is that the disturbing
events cannot be distinguished from the faults in a logical sense. However, it might
still be possible to quantitatively decouple the disturbing events from the faults using
model-based techniques.
61
Chapter 4: System Analysis and Fault Modelling
sensor configuration on the centrifugal pump. In the first subsection the division of the
system into useful components is described. In the second subsection each of these
components are then analysed for their faults and fault effects. This is done by using
the FMEA. The result of this analysis is a set of possible faults in the system. 7 of
these faults are chosen for further investigation. Using this subset, different sensor con-
figurations are analysed, showing their capability in fault detection and isolation, when
signal-based fault detection methods are used.
The centrifugal pump under investigation in this section should be seen as a general
centrifugal pump, meaning that the faults identified are general centrifugal pump faults.
Therefore in real application only a subset of these faults will occur in practice. Even
though no particular type of pump is chosen in the investigation, a CR5-10 Grundfos
pump is used as an example, whenever it can illustrate the presented ideas.
In the middle of Fig. 4.8 the centrifugal pump is shown, and each component, iden-
tified in Fig. 4.7, is shown in separate subfigures. On the top left corner the induction
motor, driving the pump, is shown. In the physical structure diagram this motor is again
divided into an electrical and a mechanical component.
Below the motor, the centrifugal pump is shown. The centrifugal pump is also di-
vided into two components in the physical structure diagram, namely a hydraulic and a
mechanical component. The hydraulic component covers the parts directly involved in
the energy transfer from mechanical to hydraulic energy, and the mechanical component
covers the remeaning parts. The middel figure to the left is a zoom of the hydraulic parts.
Here the impeller and the guide vanes are seen.
On the top to the right the shaft is shown. The CR5-10 Grundfos pump has a short
shaft connection. In other pumps this can be longer and often positioned horisontal and
not vertical as in the CR5-10 case. The two last figures show the inlet and outlet part of
the pump. In some pump applications the inlet part can be equiped with a filter to avoid
62
Section 4.3: Pump Applications
Electrical parts of
Pump Motor
the motor
Mechanical parts
of the motor
Mechanical parts
Hydraulics of the centrifugal
pump
Hydraulic parts of
the centrigfugal
pump
Inlet parts
Outlet parts
Figure 4.7: Physical structure diagram of the centrifugal pump. First the pump is divided
into three components; the motor, the shaft, and the hydraulics. These components are
again divided into 7 components.
63
Chapter 4: System Analysis and Fault Modelling
Outlet part
Inlet part
Figure 4.8: A real pump example of the component chosen for the FMEA. The pump is
a CR5-10 Grundfos pump drived by a 1.5 [KW] induction motor.
events, and their effects on the components, the propagation matrices are presented in
Appendix A.
64
Section 4.3: Pump Applications
Load torque
Vib. Temp.
3 Vib. Temp.
Mech. part
of the motor 6
Current Temp.
Mech. part
Vib. of the pump
1 2 4
Sup. voltage Electrical Vib.
part of the Mech. Shaft 5
motor Motor torque dynamics Imp. speed
Hyd. part of
Impeller head 9
the pump
leakage flow
Pressure press. diff.
Motor speed
Inlet flow
difference
8
Outlet of the
Inlet flow pump Outlet flow
7
Inlet of the
Inlet pressure pump Pressure drop at the inlet
Figure 4.9: The functional connection between the components used in the FMEA of
the pump system.
Comp. Name
c1 Electrical part of the induction motor
c2 Mechanical dynamics
c3 Mechanical part of the induction motor
c4 Shaft
c5 Hydraulic part of the centrifugal pump
c6 Mechanical part of the centrifugal pump
c7 Inlet part of the pump
c8 Outlet part of the pump
c9 Pressure difference
65
Chapter 4: System Analysis and Fault Modelling
c2 : Mechanical dynamics
This component contains the mass of all rotating parts in the pump. It is introduced
in the functional model to cover the convertion from torque to speed. As it is not a
physical component, but a functional signal transformation, no faults are identified in
the component.
c4 : Shaft mechanics
This component contains the shaft and the shaft connection attaching the motor and
pump shaft. The functionality of the component is to transfer the torque on the motor
shaft to torque on the pump shaft. The faults identified on the component are,
fsh1 Broken shaft.
fsh2 Misalignment between the motor and pump.
fsh3 Bend shaft.
The main effect of these faults are mechanical vibrations and, in the case of the last two
faults, torque oscillations.
66
Section 4.3: Pump Applications
The main effects of these faults are changes in the value of pressure and the load torque
generated by the impeller at a given flow. Moreover some of the faults can induce pres-
sure oscillations. These pressure oscillations can be either harmonics of the rotational
frequency, or noise like signals covering a larger frequency span.
Beside the faults described above the following set of disturbing events can affect
the component.
67
Chapter 4: System Analysis and Fault Modelling
The first two of these faults are connected to the cooling of the pump, and therefore the
main effect of these faults is an increased temperature of the pump. The main effects of
the last four faults are mechanical vibrations and leakages from the pump casing.
Beside the faults described above the following set of disturbing events can affect
the component.
dmp1 Decreased flow through the pump.
dmp2 Increased flow through the pump.
68
Section 4.3: Pump Applications
Sensor components
This work is not concerned with sensor faults, even though it is an important field.
Instead sensors are seen as components, which are able to measure special effects on the
system. Hence, they collect the end-effects used in the sensor analysis.
The list of possible sensors is long, some of the most important, with respect to the
centrifugal pump applications, are the following,
Current sensors.
Voltage sensors.
Vibration sensors on the stator and/or the pump mechanics.
Pressure difference sensor (between outlet and inlet).
Absolute pressure sensor at the inlet and/or the outlet of the pump.
Flow sensor at the inlet and/or the outlet of the pump.
Temperature sensors inside the stator, bearings and/or seals.
Speed sensors, on the motor and/or the pump shaft.
Of these especially the pressure difference sensor is often used for control purposes.
Likewise, if the speed of the motor is controlable normally expressions of the currents
and the voltages of the motor are also available. If the system is equipped with a surveil-
lance system, a subset of all the sensor information can be available. Unfortunately, this
is only the case in very few centrifugal pump applications.
69
Chapter 4: System Analysis and Fault Modelling
As neither risk assessments nor frequencies are identified for the faults on the cen-
trifugal pump these cannot be used as guidelines for choosing faults for further investa-
gation. Instead a set of fault is chosen using the following criteria:
The first of these criteria is included to insure the relavance of the faults from an indus-
trial point of view. The second criterion is included to make it possible to simulate the
fault behaviour in real time on a test-bench.
Using these criteria the following set of faults is chosen for futher investagation,
Of these faults fi1 , fmp1 , fip1 , fop1 , and fip2 are input faults meaning that they are
caused by unsuitable operation of the application in which the pump is placed. These
type of faults are also called external faults. Even though these faults are not dirrectly
connected to the components of the pump, they are very important to detect. The reason
is that dry running will destroy the bearings of the pump in a few seconds, and too
low inlet pressure will cause cavitation inside the pump, which again will destroy the
impeller over time.
Beside these faults of course all disturbing events affecting the system must be taken
into account in the following analysis. The disturbing events are,
70
Section 4.3: Pump Applications
Collecting the chosen fault and disturbing events the following logical fault vector is
obtained,
T
fe = f T dT (4.16)
where,
f = fe1 fh1 fh2 fh3 fm1 fin1 fin2
(4.17)
d = de1 de2 de3 de4 de5 dh1 dh2 dh3 .
In the next section the effects of this fault vector on different measureable signals are
analysed.
Table 4.4: The component description of the centrifugal pump. The structure of each of
the logical matrices is given in Appendix A.
Having a model on this form Algorithm 4.2.1 and Theorem 4.2.1 can be used to de-
rived the connection from the faults and disturbing events, to any effect vector e in the
system. Therefore, by identifing all measureable effects of interest it is possible to estab-
lish a connection between faults and disturbing events, and a subset of the measurable
effects. This connection can then be used to evaluate the usability of the given sensor
configuration, when the design of signal-based fault detection schemes is considered.
In the previous section a list of sensors used on centrifugal pumps is presented.
Moreover, it is argued that some of these sensors are only used in special applications.
The most frequently used sensors are the electrical sensors and the pressure difference
71
Chapter 4: System Analysis and Fault Modelling
sensor. Therefore these should attend special attention when developing intelligent FDI
algorithms. The electrical and pressure sensors are often used for control purposes in
hydraulic applications, and are therefore often available for other purposes too. This
means that the cost of implementing a supervision system is reduced considerably by
using only these sensors as input to the FDI algorithms.
Beside the sensors just mentioned, the flow sensor is considered. Flow sensors are
normally expensive, but by using the newest micro technology it is possible to reduce
the cost considerably. This means that this sensor will become interesting in even small
centrifugal pump applications. Also a sensor measuring the impeller eye pressure is
considered. The flow sensor and the impeller eye pressure sensor are considered to be
additional sensors and are therefore increasing the cost of implementing the supervision
system.
To summarize; the effects seen in the following sensors are analysed using the FPA,
Current sensors.
Voltage sensors.
Flow sensor.
The results of the FPA using these sensors are presented in the following, where each of
the logical matrices are obtained by using Algorithm 4.2.1.
The effects measurable using the current and voltage sensors are all found in component
c1 Electrical Part of the Motor. The measurable effects on this component are,
72
Section 4.3: Pump Applications
The connection between these effects and the faults and disturbing events in the system
is given by the following logical equation,
eem,i1 1 0 0 1 1 0 0 0 1 0 0 1 0 1 0
eem,i2 0 1 1 0 0 1 1 0 0 1 1 0 0 0 1
eem,i3
1 1 0 0 1 0 0 1 0 0 0 0 0 0 0 f
eem,i4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 d
eem,v1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
eem,v2 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
(4.18)
where f is the fault vector and d is disturbing event vector, both defined in (4.17).
where f is the fault vector and d is disturbing event vector, both defined in (4.17).
73
Chapter 4: System Analysis and Fault Modelling
The connection between these effects, and the faults and disturbing events in the system
is given by the following logical equation,
eq1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
eq2 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 f
d
eq3 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
(4.20)
where f is the fault vector and d is disturbing event vector, both defined in (4.17).
The connection between these effects and the faults and disturbing events in the system
is given by the following logical equation,
eeh1 0 0 1 0 1 0 1 0 0 0 0 0 1 0 0
eeh2 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 f
d
eeh3 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0
(4.21)
74
Section 4.3: Pump Applications
connections between faults and effects were established for each of the sensors in the
previous section.
The robustness of the different sensor configurations could be analysed using Theo-
rem 4.2.2, but by looking at the description of the disturbing events in Section 4.3.3 the
following mutually exclusive expressions of the disturbing events are recognized,
as de2 , de4 and dh2 are increased voltage, frequency and flow respectively, and de3 , de5
and dh1 are decreased voltage, frequency and flow. These dependencies are taken into
account using Corollary 4.2.1. When the robustness properties are establised the pos-
sibilities for fault identification of the detectable fault can be analysed, using Theorem
4.2.3.
The results are presented in two logical vectors Rr and Ri , where Rr contains
the results of the robustness analysis, and Ri contains the results of the identificability
analysis. Such that,
where fj is the j th component in the fault vector f and rj is the j th component in either
Rr or Ri dependent on which property is analysed. The fault vector f is given by,
f= fe1 fh1 fh2 fh3 fm1 fin1 fin2
where each fj is described in Section 4.3.3. Three sensor configurations are considered
in the following, these are,
1. Sensors measuring the electrical quantities, e.i. the voltage and current measure-
ments.
2. Sensors measuring the electrical quantities and the pressure difference across the
pump.
3. Sensors measuring the electrical quantities, the pressure difference and the volume
flow.
4. Sensors measuring the electrical quantities, the pressure difference, the volume
flow and the impeller eye pressure.
The results of the analysis of these different sensor configurations are presented in the
following.
75
Chapter 4: System Analysis and Fault Modelling
Electrical measurements
Firstly, only the effects in the electrical measurements are considered, e.i. the effects
measurable using the current and voltage sensors. The connection from the faults and
disturbing events, to their effect on the electrical quantities is given by (4.18). This
equation is on the form,
f
e Af Ad
d
meaning that Theorem 4.2.2 and Corollary 4.2.1 can be used to establish the logical
robust fault detection possibilities. The result of this analysis is shown below,
Rr = 0 0 0 0 0 0 0 . (4.22)
From Rr it is seen that it is not possible to distinguish any of the faults from possible
logical combinations of disturbing events. This means that non of these effects can be
used in a robust signal-based fault detection scheme.
76
Section 4.4: Detection Algorithm for the Centrifugal Pump
Comparing these results, with the results of the analysis where only the electrical mea-
surements and the pressure sensor are considered, it is seen that no addition information
is added using the flow sensor. This is in fact true when only logical combinations are
considered. But, as it will be shown in Chapter 6, The flow sensor can be used for dis-
turbance decoubling, when model-based methods are used. Hereby, it becomes possible
to detect the fault fh3 corresponding to increased leakage flow inside the pump.
Here it is seen that only the faults fe1 and fh1 corresponding to inter-turn short circuit
and leakage flow are undetectable. To check the possibilities for identification of the
the 5 detectable faults, the effects not corrupted by disturbing events are analysed using
Theorem 4.2.3. The logical expression of the faults is given by the following expression.
eem,i4 0 0 0 0 0 0 0
eem,v1 0 0 0 0 0 0 0
edh3 0 0 0 0 0 1 0
edh4
0 1 0 0 1 0 0 f .
edh6 0 0 0 0 0 1 0
eq3 0 0 0 0 0 1 0
eeh2 0 0 0 0 0 1 0
eeh3 0 0 1 0 0 0 1
In this expression the effects, which can be corrupted by disturbing events, are removed.
The result of the detectability analysis is shown below,
Ri = x 0 0 x 0 1 0 (4.27)
where x corresponds to the fault not robust detectable. Here it is seen that only one fault
is distinguishable from the other faults, when all the effects corrupted by disturbing
events are removed.
77
Chapter 4: System Analysis and Fault Modelling
78
Section 4.4: Detection Algorithm for the Centrifugal Pump
In this expression the same notation as in Chapter 3 is used. The effect edh3 corresponds
to zero difference pressure, when the pump is running. Finally, the effect edh4 corre-
sponds to low freqency harmonic oscillations in the pressure measurements. Here low
frequency components correspond to frequencies larger than 0 [Hz] and up to 2-4 times
the supply voltage frequency.
tank
Hp
Vl
V1 V2
pump V3
Vc
pipe Qp
Te
shaft
M
motor
Figure 4.10: Sketch of the test-bench. The measurements are the electrical quantities,
the pressure difference Hp delivered by the pump and the volume flow through the pump
Qp .
In the tank and pipe system, connected to this pump, the valve V1 is used to model
disturbances in the system. The inter-turn short circuit is simulated by shorting windings
in phase a in the costimized designed stator particular developed for this purpose. Dry
running is simulated by closing V2 and opening V3 , and rub impact is simulated by
adding an extra force on the shaft. During the test, presented here, this is done by
mounting twist on the shaft, which rubs against the mechanical connection between the
pump and motor. Hereby an oscilating force, similar to the one expected in a pump
during fault fm1 , is added. Finally, clogging inside the pump can be simulated by the
closing valve Vc . However, this valve simulates clogging of for example an inlet filter,
79
Chapter 4: System Analysis and Fault Modelling
Table 4.5: Summing of the test results. Here, the faults denoted inter-turn, Dry-running,
and rub-impact corresponds to fe1 , fm1 , and fin1 respectively. Likewise, and increase in
is , and h corresponds to eem,i3 , and edh4 respectively. Finally, when h approximate
zero it corresponds to edh3 .
Normal Inter-turn Rub-impact Dry-running
is 2.7366 5.2066 4.6263 2.0450
is 0.0037 0.4954 0.0141 0.0041
h 2.2178 2.2561 1.6402 0.1054
h 0.0003 0.0006 0.0067 0.0010
and not clogging in one of the channels in the impeller, which was assumed in the logical
analysis. Therefore, results from this test are not considered here.
The measurable effects considered in these tests are affecting the current and the
pressure measurements, therefore only these will be analysed in the following test re-
sults. To evaluate the robustness of the approach, signals obtained on the pump at con-
stant speed and at different positions of valve V1 are analysed. The different valve po-
sitions simulates the no fault condition at different hydraulic loads. Results from this
test are shown in Fig. 4.11. From these test results it is obvious that the DC-level of
the considered signals are not usable for fault detection. This was also predicted by the
FPA-analysis performed in Section 4.4.1.
Figs. 4.12, 4.13, and 4.14 depitch the current and pressure signals when the pump
is exposed to the three faults fe1 , fm1 , and fin1 denoting inter-trun short circuit, rub
impact, and dry running respectively. The results of these tests are summarized in Table
4.5. In the evaluation of the results is and h are used as measurements of the end-
effects eem,i1 and eh4 respectively. The end-effect edh3 is assumed triggered when
h 0.
Considering the results presented in Table 4.5 it is seen that is is increased consid-
erably in the case of the inter-turn fault fe1 and the rub-impact fault fm1 . Comparing
these results with the decision logic in (4.29), and rembering that is is a measure of
the end-effect eem,i1 it fits perfectly. Likewise, by using h as a measure of the edh4
it is seen that the rub-impact faults fm1 is the only fault which increases h . This also
fits the results of (4.29) perfectly. Finally, the only fault forcing h close to zero is the
dry-running fault fin1 . As h 0 is considered a measure of edh1 this also fits the
results of (4.29).
4.5 Discussion
In the literature it is described how Failure Mode and Effect Analysis (FMEA) and Fault
Propagation Analysis (FPA) are used in the design of Fault Tolerant Control (FTC).
However, these tools are general analysis tools, and could as well be used in the design
80
Section 4.5: Discussion
Current [A]
2
= 2.7366
1 = 0.0037
0
0 2 4 6 8 10 12 14 16
Time [sec]
1 = 2.2178
= 0.0003
0
0 2 4 6 8 10 12 14 16
Time [sec]
Figure 4.11: Results obtained when running the pump at constant speed and different
positions of valve V1 . The change in valve position simulates different load condition of
the pump.
6
Current [A]
= 5.2066
2
= 0.4954
0
0 5 10 15 20
1 = 2.2561
= 0.0006
0
0 5 10 15 20
Time [sec]
Figure 4.12: Results obtained when introducing an inter-turn short circuit in phase a of
the induction motor stator. The fault is introduced at time 6.05 [sec] and removed at time
5.75 [sec]. The mean and variance of the two presented signals are calculted from data
between the two indicator lines. During the test the pump is running at constant speed
and with valve V1 fixed at a constant position.
81
Chapter 4: System Analysis and Fault Modelling
Current [A]
4
= 4.6263
3
= 0.0141
2
0 1 2 3 4 5 6 7 8 9
2
Pressure [bar]
1.5
= 1.6402
1
= 0.0067
0.5
0 1 2 3 4 5 6 7 8 9
Time [sec]
Figure 4.13: Results obtained when introducing a rub-impact fault on the pump. Here,
the pump is affected by the fault during the whole data series. The mean and variance of
the two presented signals are calculated from data between the two indicator lines. Dur-
ing the test the pump is running at constant speed and with valve V1 fixed at a constant
position.
of Fault Detection and Identification (FDI) alogrithms. In this chapter these algorithms
are used as analysis and design tools in the design of signal-based FDI algorithms.
In the first part of the chapter some theoretical considerations on using FPA in the
design of signal-based FDI are considered. It is well known that some workarounds are
necessary in the FPA when loops occur in the system model. Normally, this is handled
by cutting the loops and then treat the cutted effects as additional faults in the system. In
this chapter an algorithm for identifing the optimal cuts in the loops is developed. From
the result obtained using this algorithm it is easy to find the connection between the
faults in the system and any set of end-effects. This can be done by using a theorem also
prestented in this chapter. Hereby the step of cutting loops in FPA is fully automated,
meaning that the only manual work necessary in the FPA is to set up the event model of
the system. The developed algorithm can be used in the design of FTC as well as FDI.
One of the main concerns in the design of FDI algorithms is how to handle dis-
turbances in the system. This is necessary to avoid generating fault alarms. To treat
this problem in the frame work of the FPA, it is proposed to define a set of disturbing
events. These disturbing events are treated as faults in the FPA analysis, meaning that
the connection between faults as well as disturbing events can be established using the
automated FPA. When this connection is established the connection between disturbing
events and end-effect is used to identify those of the end-effects, which will be corrupted
by disturbing events. Hereby it is possible to find the end-effects, which can be used for
82
Section 4.5: Discussion
Current [A]
3
2
= 2.0450
1
= 0.0041
0
0 5 10 15 20 25
1.5
= 0.1054
1 = 0.0010
0.5
0
0 5 10 15 20 25
Time [sec]
Figure 4.14: Results obtained when introducing a dry running fault at the test setup.
The fault is introduced at time 3.5 [sec] and removed at time 10 [sec]. The mean and
variance of the two presented signals are calculated from data between the two indicator
lines. During the test the pump is running at constant speed and with valve V1 fixed at a
constant position.
fault detection in a robust manner. A theorem for doing this is developed in the chapter.
In the second part of the chapter the FMEA and FPA are used in an analysis of
a centrifugal pump. First the FMEA is used to identify all faults expected to happen
in a centrifugal pump. This result can be used when a fault detection scheme should
be developed, as it contains information about the faults, which should be expected in
a centrifugal pump. In the presented analysis a list of possible disturbing events are
also included. The results of the FMEA are used in the FPA to analyse different set
of measurable end-effects for their fault detection capabilities. The analysed end-effects
are all measurable by conventional sensors, meaning that only voltage, current, pressure,
and flow sensors are considered. This analysis shows that the sensor configurations
analysed have pure detection capabilities, when all identified disturbing events are taken
into account. However, by relaxing the number of disturbing events a robust signal based
detection algorithm is developed, using only current and pressure measurements. This
algorithm is tested on a test-bench, where it is shown to work as expected.
83
Chapter 4: System Analysis and Fault Modelling
84
Chapter 5
Stator faults are according to (Kliman et al., 1996) the most common electrical faults
in electrical motors. Moreover according to (Bonnett and Soukup, 1992) most of these
faults start as an inter-turn short circuit in one of the stator coils. The increased heat due
to this short circuit will eventually cause turn to turn or turn to ground faults, and finally
lead to a breakdown of the stator, (Wiedenbrug et al., 2003) and references included. The
time, from an inter-turn short circuit has occurred to breakdown of the stator, can be very
short. In (Gerada et al., 2004) it is argued that the time from an inter-turn short circuit
has occured to the temperature in the short circuit exceed the breakdown temperature of
the insulation can be as small as 1 to 2 [sec].
Inter-turn short circuits are caused by several different influences on the stator. For
example mechanical stress during assembling or during operation can create scratches
in the insulation, which again can cause short circuits. If the motor is placed in wet
environment, moisture can cause flow of current from scratch to scratch, which can
make a hot spot and thereby destroy the insulation. Moreover, if the motor is supplied
with a PWM voltage source, partial discharges due to very high amplitude alternating
voltage between the turns can degrade the insulation over time and cause a short circuits.
In the literature different approaches are proposed for detection of inter-turn short
circuits. In (Cruz and Cardoso, 2001) the stator currents are transformed using the Park
transformation. Second order harmonics in the length of the transformed current vector
is then used for fault detection. In (Cash, M. A. et al., 1997; Garca et al., 2004) oscilla-
tions in the voltage between the line neutral and the star point of the motor are used as
a fault indicators. This is also shown in (Tallam et al., 2002) using a model of a faulty
motor. In (Lee et al., 2003) estimation of the negative impedance of the motor is used
as a fault indicator, and in (Arkan et al., 2001) the negative sequence current is used for
the same purpose. In (Briz et al., 2003) high frequency voltage injection in the supply
85
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
voltage is utilized to create a response on the motor current. This response contains
information of the inter-turn short circuit fault.
In this chapter a model-based approach is proposed. The proposed approach is based
on a model of the induction motor including an inter-turn fault in the stator. Different
approaches for modelling inter-turn short circuits in the stator windings are found in the
literature. In (Joksimovic and Penman, 2000) a higher order model is used. This model
is an extension of the model presented in (Luo et al., 1995). This type of model is used
for simulating higher order effects in the motor, but the obtained model is of high order.
The inter-turn short circuit fault has its main harmonics in the lower frequency range.
Therefore observers designed on the basis of this type of model will be of unnecessary
high order for this kind of fault. In (Williamson and Mirzoian, 1985) a steady state
model of both inter-turn and turn-turn faults in an induction motor is developed using a
low order model. In (Tallam et al., 2002) a transient model of the same order as the one
presented in (Williamson and Mirzoian, 1985) is developed. This model describes an
Y-connected induction motor with an inter-turn short circuit in phase a.
In this chapter an adaptive observer is proposed for estimation of the inter-turn short
circuit fault. Theoretical considerations on adaptive observers can for example be found
in (Besancon, 2000; Rajamani and Hedrick, 1995; Cho and Rajamani, 1997). Based on
these contributions a new observer scheme is proposed, specially designed for handling
bi-linear systems. The observer is formulated in general terms, hence is usable in other
applications. The proposed observer is capable of simultaneously estimating the speed
of the motor, the amount of turns involved in the short circuit, and an expression of the
current in the short circuit. The observer is based on a model, developed particular for
this purpose. This model is based on the same ideas as the model described in (Tallam
et al., 2002). However, the model developed in this chapter is valid for both Y- and
-connected induction motors, and does includes both inter-turn and turn-turn short
circuits. Moreover, the model has a more useful structure compared to (Tallam et al.,
2002). Using three copies of the designed adaptive observer the phase affected by the
inter-turn short circuit is identified using an approach described in (Zhang, 2000).
As a model-based approach for fault estimation is proposed in this chapter, the chap-
ter starts by deriving a model of the induction motor with an inter-turn short circuit in
Section 5.1. This model is in Section 5.2 used in the design of the proposed adaptive
observer. In Section 5.3 test results from tests on a customized designed motor are pre-
sented. Finally concluding remarks end the chapter.
86
Section 5.1: Model of the Stator Short Circuit
A turn-turn short circuit denotes a short circuit between windings in two different
phases of the stator, see Fig. 5.1. Here a short circuit between phase a and b in a
Y-connected and a -connected stator is shown.
isb
b +
+ -
if c
vsb vsc isc vsb
isa a
-
- if b +
+ vsa - i sa
- isb
vsc c
a
+ isc + vsa -
(a) Short circuit between turns in phase a and (b) Short circuit between turns in phase a and
b in a Y-connected stator. b in a -connected stator.
Figure 5.1: Simplified electrical diagram of a three phase Y-connected and -connected
stator with a turn-turn short circuit between phase a and b.
An inter-turn fault is, in contrast with the turn-turn fault, a short circuit between
windings in the same phase coil. However, an inter-turn fault can be treated as a special
case of the turn-turn fault, as it can be modelled by assuming that no turns, of for ex-
ample phase b, are involved in the short circuit. However, it can be argued that this is a
rather limited model assumption for the inter-turn fault, because the short circuit always
must be connected to the end point of the phase coil in this case. But, if the electrical
circuit is assumed linear, all short circuits in a coil can be represented by a short circuit
connected to the end point of the given coil. This new short circuit must of cause have
the same amount of turns as the real short circuit.
In the following, a model of an induction motor, including a turn-turn short circuit
between phase a and b, is developed. The model is developed under the assumption that
the short circuit does not affect the overall angular position of the coil in the motor.
First the Y-connected motor is considered. Setting up the mesh equations for the Y-
connected motor shown in Fig. 5.1(a) and rearranging these equations, a model de-
scribing a motor with a short circuit between phase a and b is found. Using the matrix
87
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
notation presented in Chapter 3 this model is given by the following set of equations,
d sabc
vsabc = rs (isabc if ) + (5.1a)
dt
dif
T vsabc = rf if + lf (5.1b)
dt
d rabc
0 = rr irabc + (5.1c)
dt
sabc = ls (isabc if ) + lm ()irabc (5.1d)
rabc = lr irabc + lm ()(isabc if ) , (5.1e)
where (5.1a) and (5.1d) describe the voltages and the flux linkages in each of the stator
phases, (5.1c) and (5.1e) describe the voltages and the flux linkages in each of the rotor
phases, and finally (5.1b) describes the current in the short circuit. In these equations
vsabc contains the voltages across each stator phase, isabc is the current running into
each stator phase, and if is the current in the short circuit. The matrices rs , rr , ls , lr ,
and lm () have the same form as in the case of a motor with no faults. These matrices
are given in Section 3.2.1.
The vector in (5.1a) to (5.1e) represents the position and the amount of turns in the
short circuit. The vector is, in the case of a short circuit between phase a and b, given by
T
= a b 0 , (5.2)
where a is the amount of turns affected in phase a, and b is the amount of turns affected
in phase b by the short circuit. The inductor and the resistor in (5.1b) are given by
lf = (a (1 a ) + b (1 b )) lls , rf = (a (1 a ) + b (1 b )) rs + ri (5.3)
where rs is the stator resistance, lls is the leakage inductance of the stator, and ri is the
resistance in the insulation break. ri = means that no short circuit has occurred and
ri 6= means that a leakage current is flowing. The evolution from ri = to ri = 0
is very fast in most insulating materials, meaning that the value of ri can be assumed to
equal either or 0 in most cases.
88
Section 5.1: Model of the Stator Short Circuit
This model has the same structure as the one modelling the Y-connected motor. More-
over the parameters rs , rr , rf , ls , lr , lm and lf in this model have the same values as in
the model of the Y-connected motor. The only difference is the vector modelling the
amount of turns involved in the short circuit, which in this case is given by
T
= a b 0 . (5.5)
89
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
generates air gab flux. Rewriting (5.6) and (5.7) using the same procedure as described
in Section 3.2, and introducing the current i0sdq0 the induction motor model becomes,
di0sdq
L0sdt = (Rs + R0r )i0sdq + (R0r zp r JL0m )imdq + vsdq
di0
lls dts0 = rs i0s0 + vs0
dimdq (5.8)
L0m dt = R0r i0sdq (Rr0 zp r JL0m )imdq
di
lf dtf = rf if + T T1 dq0 vsdq0 ,
where
R0r = Lm L1 1
r Rr Lr Lm L0s = Ls Lm L1
r Lm L0m = Lm L1
r Lm ,
90
Section 5.1: Model of the Stator Short Circuit
From (5.9e) it is seen that i0s0 = 31 (a b )if as it0 = 0 in the Y-connected case.
Using this expression in (5.9b), and using the obtained expression to eliminate vt0 v0
in (5.9d) a new expression of the short circuit current if is found. This expression is
given by
dif
Lf dt = Rf if + T T1
dq vtdq , (5.10)
where T1 1
dq is a matrix consisting of the two first columns of Tdq0 , and Lf and Rf are
scalars and are given by
Lf = lf + 31 (a b )2 lls Rf = rf + 31 (a b )2 rs .
The final model of the Y-connected induction motor then becomes
di0
L0s dt sdq
= (Rs + R0r )i0sdq + (R0r zp r JL0m )imdq + vtdq
di
L0m dtmdq
= R0r i0sdq (R0r zp r JL0m )imdq
di
(5.11)
Lf dtf = Rf if + T T1 dq vtdq
itdq = i0sdq + Tdq if ,
T
where = a b 0 .
91
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
In the -connected case the current i0s0 = is0 13 (a b )if . From (5.12b) it is seen
that i0s0 0 as t , meaning that is0 31 (a b )if as t . This shows that
the circulating current in the -connected motor will be proportional to the current in
the short circuit if .
The final model of the -connected induction motor with a short circuit fault be-
comes
di0
L0s dt sdq
= (Rs + R0r )i0sdq + (R0r zp r JL0m )imdq + Bv vtdq
di
L0m dt mdq
= R0r i0sdq (R0r zp r JL0m )imdq
di
(5.13)
lf dtf = rf if + T T1 dq Bv vtdq
itdq = Ci i0sdq + Ci Tdq if ,
T
where = a b 0 in the -connected case, and T1
dq consists of the two first
1
columns of Tdq0 as in the Y-connected case.
lf m ()
Te = zp iTs ir , (5.14)
where is is the current in the stator windings and ir is the current in the rotor. In the case
of a stator with faults, the current in each part of the faulty windings must be defined.
In the case of a stator with a single turn-turn short circuit between phase a and b, the
currents is and ir are given by
T
is = isa isa sign( 1 )if isb isb sign( 2 )if isc
T
ir = ira irb irc ,
where 1 and 2 are the first and second elements in respectively. The matrix
lf m (zp r ) is the mutual inductance matrix between the stator and rotor. lf m (zp r )
92
Section 5.2: An Adaptive Observer for Inter-turn Fault Detection
T lm (zp r )
Te = zp (isabc if ) irabc , (5.15)
where is given by (5.2) and (5.5) in the Y- and -connected cases respectively. lm ()
has the same structure as in the no fault case, and is given in Section 3.2.1.
Transforming the torque expression in (5.15) using the transformation Tdq0 and us-
ing that Lm imdq = Lr irdq + Lm isdq , the following torque expression is obtained,
3
Te = zp L0m imd i0sq imq i0sd
2
where L0m = L2m /Lr , meaning that L0m is the diagonal element of L0m .
Remark 5.1.1 By examining the electrical model of an induction motor with an inter-
turn short circuit (5.11) or (5.13) and the torque expression given above, it can be seen
that no torque ripples should be expected, if the load is constant at a given speed and
the motor is supplied with a balanced three phase sinusoidal voltage.
93
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
identical observers, which can detect faults in phase a, b and c respectively, can be used
for identification. The approach used for this is described at the end of this section and
is based on (Zhang, 2000).
In the model developed in the previous section an inter-turn short circuit in phase a
is modelled by setting b = 0 in the vector . Doing this the parameters Rf and Lf in
both the Y- and -connected motors can be expressed by
Rf = f (a )rs + ri Lf = f (a )lls ,
The only differences between the model of the Y- and -connected induction motors
are the structures of Bv , Ci , and f (a ). In the Y-connected case Bv = Ci = I and in
the -connected case Bv and Ci are given in Section 3.2.3 of Chapter 3.
The model in (5.16) represents an induction motor with an inter-turn short circuit in
the stator. However, by setting a = 0 a motor not affected by a fault can be modelled.
This is true because setting a = 0 represents a short circuit involving 0 turns, which
have the same effect on the motor performance as when no short cicuits have occured.
a
To ensure the validity of this model the term f ( a)
must be bounded when a 0. The
term is given by
a a
= ,
f (a ) a (1 aa )
where a equals 23 and 1 in the Y- and -connected cases respectively. From this expres-
a
sion it is seen that f ( a)
is bounded on the set {a : 0 a < 1}, which contains all
possible values of a .
A surprising fact when modelling the no fault case by setting a = 0 is that the fault
current if is not equal to zero in the no fault case. This is because the model expresses
94
Section 5.2: An Adaptive Observer for Inter-turn Fault Detection
the limit of the fault current when a 0, when the no fault case is modelled in this
way. The model is still correct in the no fault case, as the fault current only affects the
remaining model in the output expression (5.16b), and here the fault current is multiplied
with a which is equal to zero.
From (5.16) it is seen that the model of the induction motor contains one unknown
variable r , which represents the speed of the rotor. If this variable is modelled as an
unknown but constant parameter, an adaptive observer approach can be used for state
estimations, and thereby estimation of the fault. The design of this adaptive observer is
considered in Section 5.2.1. The design of the feedback gain in the adaptive observer
is considered in Section 5.2.2, and finally the identification of the phase affected by a
given inter-turn short circuit fault is considered in Section 5.2.3.
Remark 5.2.1 Item Number 4 in Definition 5.2.1 means that the only states, which can
be directly affected by i , are the measurable states, and that Ai (u) has the following
structure
0
Ai (u)
Ai (u) =
0
95
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
due to the structure of C = I 0 .
For systems on the form defined by Definition 5.2.1 an adaptive observer exists accord-
ing to the following proposition.
Proposition 5.2.1 For a system of the form defined in Definition 5.2.1 an adaptive ob-
server exists and has the following form,
dz
dt = A(u, )z + Bu + K(u)(y Cz)
di (5.18)
dt = i zT A0i (u)T (y Cz) i {1, , k} ,
where i > 0, i {1, , k} are design constants, A(u, ) is a copy of the system
matrix defined in Definition 5.2.1 and A0i (u) is a submatrix given by
0
Ai (u)
Ai (u) = .
0
If there exists a K(u) stabilizing the system for every D and u U, i.e. there
exists a P() fulfilling the following Matrix Inequality (MI),
T
(A(u, ) K(u)C) P() + P() (A(u, ) K(u)C) 0
(5.19)
P() 0 .
Proof: Defining the state estimation error as ez = z z, and the parameter estimation errors as
ei = i i , the dynamics of the errors become
P
ez = (A(u, ) K(u)C) ez + ki=1 Ai zei
T T
ei = i z Ai ez i {1, 2, , k} ,
where it is used that ei = i as i = 0. This error equation can be reformulated by defining
the following parameter error vector and matrix,
T
e = e1 e2 ek A (z) = A1 z A2 z Ak z .
Using this vector and matrix in the description of the error system it becomes,
ez (A(u, ) K(u)C) A (z) ez
= T
e A (z) 0 e
I 0 (A(u, ) K(u)C) A (z) ez
= , (5.20)
0 A (z)T 0 e
T
I 0 I 0
where = diag{1 , 2 , , n }, meaning that = 0. Now define e =
0 0
T T
ez eT and let the following quadratic function be a Lyapunov function candidate for the
error system (5.20),
P() 0 P() 0
V = eT e ,where 0.
0 I 0 I
96
Section 5.2: An Adaptive Observer for Inter-turn Fault Detection
Taking the derivative along the trajectory of (5.20) and recognizing that
" T #
I 0 0 A (z) P() 0 P() 0
eT + e=0
0 A (z)T 0 0 I 0 I
for all e Rn+k . This is true as the matrix in the expression is skew symmetric. Moreover,
dP()
dt
= 0 as = 0. Using these statements the derivative of the Lyapunov function candidate
becomes,
V = eTz (A(u, ) K(u)C)T P() + P() (A(u, ) K(u)C) ez .
When this transformation is used on system (5.16) it is transformed into a bilinear system
on the form,
dz
dt = (A0 + r Ar + vsd Avsd )z + Bu
(5.23)
y = Cz ,
97
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
where the input vector u = vtdq , the output vector y = itdq , and the matrices in (5.23)
are given by
" #
Rs +Rr0 rs
0 1 1 1
Ls (Rs + R0r ) L0m R0r L0s R0r L0s lls 0
0
Rs rs
Lm lls
A0 = L0m
1
Rs 0
0
0
0
0 0 lrlss 0
0 0 0 0
2
0 0 0 3lls
1
zp J zp JL0s L0m 0 0 0
2(lm +lls )
0 0 0 0
A r =
0
Avsd =
0 0 0 3Lm lls
0 0 0 0
0 0 0 0 0 0 0 2
3lls
0 0 0 0
1
L0s
L0 1
B=
0 Bv
m C = Ci I 0 0 0 .
0
In this design K(u) is the stabilizing feedback gain and 0 < is the adaptation gain.
should be chosen such that the adaptation speed is suitable. K(u) should be chosen
according to Proposition 5.2.1. The calculation of K(u) is considered in the following
section.
98
Section 5.2: An Adaptive Observer for Inter-turn Fault Detection
for all u U and D . In the induction motor case this is the same as saying that the
MI must be fulfilled for all possible values of vsd and r .
Before the design of the induction motor case is considered a general approach for
analysis and synthesis of the observer gain is given. This analysis and synthesis approach
is restricted to the subset of systems defined in Definition 5.2.1, and where A(u, ) is
given by
k
X m
X
A(u, ) = A0 + i Ai + uj Ak+j , (5.26)
i=1 j=1
T
meaning that A(u, ) is affine with respect to u and . Here Im{Ai } Ker{C} = 0 for
all i {1, , k} must be true for Definition 5.2.1 to be fulfilled. Moreover, A(u, )
must be bounded, which is the case for (5.26) whenever D and U are bounded sets.
Analysis
First the analysis of stability is considered when a candidate for K(u) is Pgiven. Assume
m
that K(u) has a structure such that it can be written as K(u) = K0 + j=1 uj Kj , i.e.
it is affine with respect to u. If this is the case the MI in (5.25) is affine with respect to u.
Moreover, assume that the sets U and D are on a form such that the set of unknown u,
in (5.25) can be descibed by a convex hull (Scherer and Weiland, 1999), see Appendix
B.2. Let this convex hull be given by
= co {0 } , 0 = (u, ) | uj {uj , uj }, i {i , i } ,
99
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
which must be fulfilled for all (u, ) 0 and all i {1, , k}. In this LMI P() is
Pk
on the form P() = P0 + i=1 i Pi .
The feasability of this LMI states that the adaptive observer in Proposition 5.2.1
is stable for the given K(u). However, if there are limits on the change rate of the
supply signals u, this rate limit can be incorporated in the stability analysis, relaxing the
condition. To see this, let u be contained in the set given by
= co {0 } , 0 = co u | uj {j , j } ,
and let P in (5.27) be an affine function of both and u. Then the stability condition
can be formulated by the following LMI (Scherer and Weiland, 1999, Prop. 2.43), see
Appendix B.2,
T
(A(u, ) K(u)C) P(, u) + P(, u) ( ) + P(0, ) P0
P(, u) 0 (5.28)
ATi Pi + Pi Ai 0 ,
where denotes a copy of the contents of the previous bracket and P(, u) is on the
Pk Pm
form P(, u) = P0 + i=1 i Pi + j=1 uj Pj+k . The LMI must be fulfilled for all
(u, ) 0 , 0 and all i {1, , k + m}.
Remark 5.2.2 The LMIs presented in (5.27) or (5.28) are not the only LMI formulation
for stability analysis of (5.18). The problem can also be transformed to the standard
formulation used in the robust control community. Doing this performance demands
defined in the frequency doman can be included in the analysis. However, this approach
is not prestented here, as it is the author oppinion that (5.27) and (5.28) are easier to
follow, due to the straightforward connection to Lyapunov stability.
Synthesis
The LMI presented in (5.27) and (5.28) can only be used for analysis, because P is a
function of D and u U. However, introducing the restriction
Pm that P is constant
over D and U, the LMI is made solvable for K(u) = K0 + j=1 uj Kj . Introducing
this restriction in (5.25) it becomes
T
(A(u, ) K(u)C) P + P (A(u, ) K(u)C) 0
(5.29)
P 0 ,
which must be fulfilled for all (u, ) 0 . Using the transformations Lj = Kj P
j {0, 1, , m}, (5.29) can be rewritten to become
T
Pk
P A0 + i=1 k Ai L0 C + +
" T #
Pm (5.30)
j=1 uj PAk+j Lj C + 0
P 0 ,
100
Section 5.2: An Adaptive Observer for Inter-turn Fault Detection
where denotes a repetition of the contents of the previous bracket. This expression is
affine in P and Lj j {1, , m}, meaning that if a solution exists for P and Lj it
can be found using the LMI (5.30). P has full rank, meaning that the transformation
Lj = Kj P is solvable for all Kj j {1, , m}.
The assumption that P() is constant for all , is the same as saying that the param-
eters in can change arbitrary fast. However, the parameters are assumed constant in
the design of the adaptation part of the observer, meaning that conservatism is introduced
by this calculation of K(u).
The only unknown parameter in the induction motor case is the speed r , meaning that
= r . Likewise, there is only one input making the system bilinear, namely vsd , i.e.
u = vsd . Based on this the convex hull for this system is given by
= co {0 } , 0 = {(vsd , r ) | vsd {v sd , v sd }, r { r , r }} ,
where v sd < 0 < v sd and r < 0 < r . In Section 5.2.1 it is argued that the induction
motor system is not observable when vsd = 0. Therefore, the MI (5.25) can never have
a solution on 0 . It is also argued that vsd only goes through zero and never stays there.
Therefore, the fraction of time where vsd = 0 tends to zero. Using this argument it is
only necessary to check stability in the set {vsd | v sd vsd v sd }\ 0. This can be
done by reformulating the MI (5.25) obtaining the following LMI,
T
(A(vsd , r ) K(vsd )C) P(r ) + P(r ) ( ) 0
T
(A(vsd , r ) K(vsd )C) P(r ) + P(r ) ( ) 0 (5.31)
P(r ) 0 ,
where denotes a copy of the contents of the previous bracket, and A(vsd , r ) = A0 +
r Ar + vsd Avsd . The matrices A0 , Ar , Avs d , B, and C in the above LMI are all
defined in Section 5.2.1. This LMI should be feasible on the set given by
for the adaptive observer to exist. For the induction motor used in the tests presented at
the end of this chapter the conservatism imposed in the synthesis is too restrictive, i.e. no
solution can be found. However, it is still possible to check stability for a given K(vsd )
without introducing conservatism in the analysis. This is done by using the analysis
approach presented by the LMI (5.28), taking limits on the change rate of the supply
signal vsd into account. Utilizing this approach the stability of the observer with a given
101
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
where denotes a copy of the contents of the previous bracket, and A(vsd , r ) = A0 +
r Ar + vsd Avsd . The matrices A0 , Ar , Avs d , B, and C in the above LMI are all
defined in Section 5.2.1. This LMI should be feasible on the set given by,
(vsd , vsd , r ) | vsd {, v sd }, vsd {, }, r { r , r } >0
for the adaptive observer to exist. In the LMI the inequalities (5.32a) to (5.32d) are
introduced to check stability of the two regions defined by 0 < vsd and vsd < 0 re-
spectively. The inequalities (5.32e) and (5.32f) are introduced to guarantee convexity
of (5.32a) and (5.32c) with respect to the parameter r and the input vsd . The analysis
approach presented by the LMI (5.32) is utilized in the design of the observer, used in
the tests prestented in Section 5.3.
Remark 5.2.3 In the above text it is argued that the observer is stable for all values of
the voltage vsd between v sd and v sd [V ] except for vsd = 0, and all speeds between
r and r [rad/sec], if the LMI (5.32) is feasiable. This is not the same as saying that it
is possible to estimate the fault and speed at zero speed, due to demands for persistence
of excitation, see the proof of Proposition 5.2.1.
Remark 5.2.4 Fault tolerant control can be obtained using the current vector i0sdq , es-
timated by the proposed observer, as input to the current controllers in a traditional
motor control system. This current is the part of the stator current producing air gab
flux. Therefore, by using this current the control is not affected by the short circuit.
This current vector is given by the two first terms of the state vector x in (5.21),
which is calculated using x = 1 (z), where is defined in (5.22).
102
Section 5.3: Test Results
indicator of the correctness of the model used in the observer design. This means that
when only one of the observer faults approximate zero it indicates that the fault modelled
by this particular observer has happened in the system. This approach can, in the case
of the induction motor, be used if three identical observers are designed, each detecting
a stator winding fault in one of the three phases. This is considered in the following.
The adaptive observer Oa in (5.24), is capable of estimating an inter-turn fault in
phase a. However, the same observer can be used to estimate a fault in phase b and c
by using another phase sequence as argument in the transformation Tdq0 . To estimate a
fault in phase b the following transformation must be used,
Using these transformations in the approach presented in (Zhang, 2000) the overall struc-
ture of the fault identification and estimation algorithm becomes as depitched in Fig. 5.2.
Here the error signals eya , eyb and eyc are used for identification of the phase affected
by a inter-turn fault, and xa , xb and xc are the estimates of the states including the
estimates of the fault current if and the fault size .
The affected phase is in this work identified by comparing the square error, of each
error signal eya , eyb and eyc , with a predefined threshold value. The square error is
calculated using eTi ei .
Remark 5.2.5 The approach used here is based on a predefined threshold value. This
threshold must be chosen as a trade-off between how small a fault can be detected and
robustness in the system. However, the problem of choosing this threshold can be over-
come by comparing the square error of the three error signals eya , eyb , and eyc . If the
levels of these error signals are alike no fault has occurred, and if a fault has occured
it is identified by the signal with the smallest square error. However, using these ap-
proaches you cannot be sure to identify the correct fault, you will just identify the fault
most likely to have occured among the modelled faults.
103
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
Real
world
Tl
r
Pump Q,H
Vs Motor is
Software is - eya
algorithm abc Observer
xa
abc Oa
is - eyb
abc Observer
xb
bca Ob
is - eyc
abc Observer
xc
cab Oc
Figure 5.2: The structure of the identification algorithm for identification and estimation
of inter-turn short circuits. Here eya , eyb and eyc are used for identification of the
affected phase, and xa , xb and xc are the estimates of the states including the estimates
of the fault current if and the fault size .
tests is a 1.5 [KW] customized Grundfos motor, supplied with a Danfoss frequency con-
verter. The speed, the three phase currents, and the three phase voltages are available at
the test setup. The voltage to the motor is controlled using a linear voltage to frequency
relation, with a voltage boost at low frequencies. All tests are preformed at supply fre-
quencies around 30 [Hz] to avoid too large short circuit currents and thereby burnout of
the motor during the tests. The tests are performed with the induction motor connected
in a -connection as it is shown Fig. 5.3. However, similar results can be found for a
Y-connected motor in (Kallese et al., 2004c).
In the first, of the two following subsections, the identification capabilities of the
proposed algorithm are tested. In the second subsection the estimation capabilities of the
adaptive observer are tested. In these tests the algorithm is tested against three different
operating conditions. These are,
Speed changes at every 1 second between 25 and 40 [Hz] and balanced supply
voltage.
104
Section 5.3: Test Results
5%
25
%
c b
25%
5%
Figure 5.3: The electrical circuit of the stator in the test setup. Two points of phases a
and b and their end points are available at the terminal box.
105
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
eA
2
0
0 5 10 15
3
eB
0
0 5 10 15
3
eC
0
0 5 10 15
time [sec]
Figure 5.4: The mean square error of the observers Oa , Ob , and Oc respectively. In
this test the speed is constant and the supply voltage is balanced, and faults are imposed
seperately in phase a and c.
the inherent imbalance in the phases, the results are comparable. This shows that the
algorithm is able to handle unbalanced supply conditions, which also was expected as
no assumption were put on the supply voltage in the design. This means that the observer
can manage any distortion of the supply voltage as long as it does not introduce too large
oscillations in the speed.
106
Section 5.4: Conclusion
eA
2
0
0 5 10 15
3
eB
0
0 5 10 15
3
eC
0
0 5 10 15
time [sec]
Figure 5.5: The mean square error of the observers Oa , Ob , and Oc respectively. In
this test the speed is variating and the supply voltage is balanced, and faults are imposed
seperately in phase a and c.
partly due to noise on the measurements, partly due to mismatch between the real motor
parameters and the motor parameters used in the observer, and partly due to the initial
imbalance between the three stator phases. This bias is repeated in each of the three
tests.
Results from the second test, presented in 5.8(a) and 5.8(b), show that the observer
is capable of estimating the wanted quantities despite of speed changes. Still it is seen
that the speed changes affect the estimated amount of turns in the short circuit. This is
because of the constant speed assumption used in the design. It is, however, still possible
to use the estimate of the fault.
From the results of the last test, presented in 5.9(a) and 5.9(b), it is seen that an
unbalanced supply of 5% is not affecting the performance of the observer.
5.4 Conclusion
An adaptive observer for simultaneous estimation of the motor states, the speed, and
the amount of turns in an inter-turn short circuit is proposed. The observer is tested
on a customized designed induction motor. The tests have shown that the observer can
estimate an inter-turn fault despite of speed changes and unbalanced supply conditions.
107
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
eA
2
0
0 5 10 15
3
eB
0
0 5 10 15
3
eC
0
0 5 10 15
time [sec]
Figure 5.6: The mean square error of the observers Oa , Ob , and Oc respectively. In this
test the speed is constant and the supply voltage is unbalanced, and faults are imposed
seperately in phase a and c.
This makes the estimation scheme usable in inverter feed induction motor drives, or in
motor applications supplied by a bad grid. Using three of these observers it is shown
that it is possible to identify the phase affected by a given inter-turn short circuit.
The adaptive observer is based on a model of an induction motor including an inter-
turn short circuit. This model is a simplification of the model presented in the start of
this chapter, as the derived model describes the motor behaviour in both the inter-turn
and the turn-turn short circuit cases. Moreover, the model describes the motor when it
is connected in a Y-connection as well as a -connection.
Comparing the approach presented here with traditional approaches, the main ad-
vance is that the obtained observer is based on a dynamic model of the system. This
means that the detection capabilities are not affected by dynamic changes in the elec-
trical system. The main drawback of the proposed approach is the need for the motor
parameters. However, in the cases where the approach is used in a frequency converter
application, this problem can be solved by parameter identification methods at start up
(Rasmussen, 1995).
The proposed adaptive observer can be used for fault-tolerant control of the induction
motor, as the impact of the inter-turn short circuit is estimated. This is so because it is
possible to obtain control in the case of an inter-turn short circuit, meaning that it is
possible to control the process, driven by the motor, to a fail-safe mode, or to reduce the
108
Section 5.4: Conclusion
6
i
f
4
Scaled current
2
6
0 2 4 6 8 10
0.3
est
real
Fraction in short circuit
0.2
0.1
0.1
0 2 4 6 8 10
time [sec]
(a) The top figure shows the estimation of the scaled current a if and
the bottom figure shows the estimated and real amount of windings af-
fected by the short circuit.
220 r,est
200 r
speed [rad/sec]
180
160
140
120
100
0 2 4 6 8 10
30
e
20
speed error [rad/sec]
10
10
20
30
0 2 4 6 8 10
time [sec]
(b) The top figure shows the estimated and the measured speed and
the bottom figure shows the error between the estimated and measured
speed e .
Figure 5.7: The results from tests of estimation capabilities of the adaptive observer. In
this test the speed is constant and the supply voltage is balanced.
109
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
8
i
6 f
Scaled current
2
0
2
4
6
8
0 2 4 6 8 10
0.3
est
real
Fraction in short circuit
0.2
0.1
0.1
0 2 4 6 8 10
time [sec]
(a) The top figure shows the estimation of the scaled current a if and
the bottom figure shows the estimated and real amount of windings af-
fected by the short circuit.
350
r,est
r
300
speed [rad/sec]
250
200
150
0 2 4 6 8 10
30
e
20
speed error [rad/sec]
10
10
20
30
0 2 4 6 8 10
time [sec]
(b) The top figure shows the estimated and the measured speed and
the bottom figure shows the error between the estimated and measured
speed e .
Figure 5.8: The results from tests of estimation capabilities of the adaptive observer. In
this test the speed is variating and the supply is balanced.
110
Section 5.4: Conclusion
6
i
f
4
Scaled current
2
6
0 2 4 6 8 10
0.3
est
real
Fraction in short circuit
0.2
0.1
0.1
0 2 4 6 8 10
time [sec]
(a) The top figure shows the estimation of the scaled current a if and
the bottom figure shows the estimated and real amount of windings af-
fected by the short circuit.
220 r,est
r
200
speed [rad/sec]
180
160
140
120
100
80
0 2 4 6 8 10
30
e
20
speed error [rad/sec]
10
10
20
30
0 2 4 6 8 10
time [sec]
(b) The top figure shows the estimated and the measured speed and
the bottom figure shows the error between the estimated and measured
speed e .
Figure 5.9: The results from tests of estimation capabilities of the adaptive observer. In
this test the speed is constant and the supply voltage is unbalanced.
111
Chapter 5: A New approach for Stator Fault Detection in Induction Motors
level of the current in the short circuit, and thereby increase the time from the occurrence
of the inter-turn short circuit to a stator burnout.
112
Chapter 6
The topic of this chapter is FDI on the hydraulic and mechanical part of the centrifugal
pump. The model-based approach is used for this purpose, meaning that a set of residual
generator are developed, each based on the centrifugal pump model presented in Chapter
3. The centrifugal pump model is highly nonlinear, which is why methods based on
a linearization of the model will, in general, fail to work on a larger area around the
operating point of the linearization. Therefore nonlinear methods should be considered,
when the operating point is changed frequently or is unknown. In a lot of applications
this is infact the case.
Beside robustness with respect to the operating point, it is important that the algo-
rithms do not depend on knowledge of the application of the pump. This means that the
developed algorithms should work in spite of the hydraulic system in which the pump
is placed. The Structural Analysis (SA) (Blanke et al., 2003; Izadi-Zamanabadi, 2001;
Izadi-Zamanabadi and Staroswiecki, 2000) is a tool designed to identify subsystems,
which are independent of the rest of the system. Therefore this tool will be chosen for
identification of subsystems, which can be used for FDI on the centrifugal pump in a
robust manner.
The common way to obtain residual generators from subsystem identified using SA
is to derive Analytical Redundance Relations (ARR) (Blanke et al., 2003). Unfortu-
nately, in general the derived ARRs are functions of the derivatives of the measurements
in the system. These are normally not known and are difficult to calculate. To overcome
this problem a novel method to derived state space realizations of the subsystems iden-
tified using SA is developed in this chapter. Using this method the obtained state space
realizations are decoupled from any unknown algebraic variables or inputs. Therefore,
the decoupling problem is solved and the only remaining problem is to design a stable
residual observer.
113
Chapter 6: A New Approach for FDI in Centrifugal Pumps
The chapter starts by presenting some preliminaries on SA in Section 6.1. Then the
state space realization, developed in this work, is described in Section 6.2. After that FDI
on the centrifugal pump is considered. This is done by presenting the model of the pump
in Section 6.3, followed by the results obtained using SA in Section 6.4. One of the re-
sults of the SA is that the system can be splitted into two cascade connected subsystems,
of which only the second is affected by the mechanical and hydraulic faults considered
in this chapter. Therefore only an observer has to be considered for the first subsys-
tem. This observer must observe the connecting variables between the two subsystems.
The design of this observer is described in Section 6.5. In Section 6.6 the design of the
residual observers are considered. These observers are based on results obtained using
SA on the second subsystem, and the realization theory developed in Section 6.2 of the
chapter. Test results obtained on an industrial test-bench, which has been particularly
developed for this purpose, are presented in Section 6.7. Finally, concluding remarks
end the chapter.
Definition 6.1.1 (System structure) (Blanke et al., 2003) A system S is defined by two
sets C and Z, where,
C is the set of constains in the system connecting the variables of the system.
The following three assumptions must hold on the constraints C in S for the SA al-
gorithms to work. Before these assumptions are presented, let Zc = Q(c) denote the
varibles constrained by c and let nc = |Q(c)| be the number of variables in c.
114
Section 6.1: Preliminaries: Structural Analysis
x1 x1 x2 x2 u
K X
u x1 x1 x2 x2
c1 1 1
c1 c2 c3 c4
c2 1 1
(a) Graph representation. c3 1 1
c4 1 1 1 1
Figure 6.1: Structural representation of the connection between set of equations denoted
constraint c1 to c4 , and the set of variables Z = {u, x1 , x1 , x2 , x2 }. Here K = {u} is
the known variable, and X = {x1 , x1 , x2 , x2 } is the set of unknown variables.
Assumption 6.1.1 (Blanke et al., 2003, Ass. 5.1) Any algebraic constraint c C defines
a manifold of dimension nc 1 in the space of the variables Q(c).
Assumption 6.1.2 (Blanke et al., 2003, Ass. 5.2) All the constraints in C are compati-
ble.
Assumption 6.1.3 (Blanke et al., 2003, Ass. 5.3) All the constraints in C are indepen-
dent.
The structural model of S, defined in Definition 6.1.1, is a model describing the connec-
tion between the variables Z and the constraints in C by a bi-partite graph. This graph is
defined in the following definition.
Definition 6.1.2 (Structural model) (Blanke et al., 2003, Sec. 5.2) The structural
model (or the structure) of the system (C, Z) is a bi-partite graph (C, Z, E) where
E C Z is the set of edges defined by:
An example of such a bi-partite graph is shown in Fig. 6.1. Both a graphical and a table
representation of the the graph are shown here. Using the bi-partite graphs defined in
Definition 6.1.2 the structure of the model equations forming the set of constraints C can
be analysed. Two important properties in this analysis are Reachability and Matching.
These are defined in the following definitions.
115
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Definition 6.1.5 (Complete matching) (Blanke et al., 2003) A matching is called com-
plete with respect to C if |M| = |C| holds. A matching is called complete with respect
to Z if |M| = |Z|.
When there exist a complete matching with respect to the unknown variables X Z
in the system, it means that, in almost all cases, these variables can be eliminated by
rewriting the system equations. To ensure this property the notation of calculability is
used. Calculability is defined in the following definition.
116
Section 6.1: Preliminaries: Structural Analysis
Table 6.1: Incident matrix of a structural graph. This structural graph is a graph of a
minimal over-constrained system, which can be a subsystem of a larger system.
The structural model presented in Table 6.1 of the system S can have, but does not
always have, one of following three properties,
In the cases where the system S fails to conform to any of the three above properties,
it can be proven that there exists a unique decomposition of S into three subsystems
(Blanke et al., 2003),
S+ = (C+ , Z+ )
S0 = (C0 , Z+ Z0 )
S = (C , Z+ Z0 Z ),
Definition 6.1.10 (Causality) (Blanke et al., 2003) A subsystem is called causal if there
exists an alternating chain from ki K to xj X for all reachable xj X and this
chain is composed of only calculable matchings.
117
Chapter 6: A New Approach for FDI in Centrifugal Pumps
where Xmin Zmin are the unknown variables contained in the set of constraints
Cmin . Additionally, Cmin C and Zmin Z.
Such a minimal over-constrained subsystem contains information enough to derived ex-
actly one residual. Therefore, if a set of minimal over-constrained subsystems is identi-
fied in a system, and the matchings, which define each of these subsystems, are causal,
a set of residual generators can be derived. When the residual generators are derived
by eliminating all unknown variables in the subsystem they are called Analytical Re-
dundant Relations (ARR). If each of these residual generators are sensitive to different
subsets of the faults affecting the system this can be used for fault identification. This is
formalized in the following two theorems.
Theorem 6.1.1 (Structural observability) (Blanke et al., 2003, The. 5.2) A necessary
and sufficient condition for system (6.1) to be structural observable is that, under deriva-
tive causality
1. all the unknown variables X are reachable from the known ones,
2. the over-constrained and the just-constrained subsystems are causal,
3. the under-constrained subsystem is empty.
For systems which are structural observable as stated in Theorem 6.1.1, and where the
over-constraint subsystem is non empty, it is always possible to identify a number of
minimal over-constraint subsystems, as defined in Definition 6.1.11.
Let one of the constraints C be corrupted by a fault. Let the unknown variables
containted in the constraint be given by X = Q(). Then the following theorem states
the conditions for the fault corrupting the constraint to be monitorable or detectable.
Theorem 6.1.2 (Monitorability) (Blanke et al., 2003, The. 5.3) Two equivalent neces-
sary conditions for a fault to be monitorable are:
(i) X is structural observable - according to Theorem 6.1.1 - in the system
(C\{}, Z),
(ii) belongs to the structurally observable over-constrained part of the system
(C, Z).
118
Section 6.2: Realization
Algorithms exist, which can decompose a system S into all possible minimal over-
constraint subsystems. For each of these minimal over-constraint subsystems the con-
nection nc = nx + 1 holds, where nc = |C| is the number of constraints and nx = |X| is
the number of unknown variables in these constraints (Izadi-Zamanabadi, 2001; Izadi-
Zamanabadi and Staroswiecki, 2000). It is possible to derive an Analytical Redundant
Relation (ARR) for each of these subsystems. Each of these ARR can be used to gen-
erate one residual, which is sensitive to a subset of the faults in the system S (Blanke
et al., 2003). The connection between the subsystem used in the derivation of a given
ARR and the faults is described by Theorem 6.1.2.
6.2 Realization
Realization denotes the task of finding a description of an over-constrained subsystem,
identified using SA, which can be used for residual generation. A straightforward way
to find such a residual generator, is to solve the set of constraints C, forming the sub-
system, for the unknown variables X in this set of constraints. The solution obtained
using this approach is an Analytical Redundant Relation (ARR) (Blanke et al., 2003).
Unfortunately in general the obtained ARRs are functions of the derivatives of the mea-
surements, which are in general not calculable due to measurement noise.
To overcome the problem with the derivatives in the ARRs, a new approach is devel-
oped here. The main idea is to find a state space description of the subsystem identified
using SA, meaning that the subsystem can be described on the form,
dz
dt = fz (z, y, u)
(6.3)
gz (y, u) = hz (z, y, u) ,
where z contains the states of the system, and u and y contain the known signals in the
system. If a state space description (6.3) exists it can be used in the development of an
observer, which enables residual generation. The decoupling of unknown inputs has not
to be considered in this observer design, as this problem is already solved using SA and
the realization techniques presented here.
Assuming that an over-constrained subsystem identified using SA is described by
the following set of equations,
xd = fx (xd , xa , u)
0 = mx (xd , xa , u)
Sx : , (6.4)
y = hx (xd , xa , u)
dxd
dt = xd
input and output signals respectively, and xd , xd , and xa are unknown signals.
119
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Let the subsystem, identified using SA, be given by (C, Z), then each equation in
(6.4) is given by a constraint in C. Likewise, the variables u, y K and xd , xd , xa X
are the known and unknown variables respectively. These variables form the set Z, i.e.
Z = K X. The graph representation of this system is shown in Table 6.2, where
Table 6.2: Incident matrix of a structural graph. This structural graph is a graph of a
minimal over-constrained system, which can be a subsystem of a larger system.
y u xd xd xc
fx 0 G I F1 F2
mx 0 M3 0 M1 M2
hx I H3 0 H1 H2
dx 0 0 I X 0
Here it is assumed that a stabilizing feedback k can be found. In the above equation r
is the set of residual outputs, u is a set of known signals defined as input in the original
system, y is the set of known signals defined as outputs in the original system, and z
contains the states of the residual filter. If the subsystem is a minimal over-constraint
subsystem the residual r becomes a scalar r. This is infact normally the case.
In the following, first an output transformation is considered, eliminating a part of
the algebraic variables xa . The remaining algebraic variables are eliminated using a
state transformation. This is considered in Section 6.2.2. Finally, the two elimination
approaches are composed into one algorithm in Section 6.2.3.
120
Section 6.2: Realization
xa denoted xa1 and a subset of the variables xd denoted xd1 are matched using mx and
hx . Then there must exist an explicite solution to these variables using mx and hx . Let
this solution be given by,
xa1 g1 (xd2 , xa2 , y, u)
= ,
xd1 g2 (xd2 , xa2 , y, u)
T T
where xa , xd are partitioned such that xa = xTa1 xTa2 , xd = xTd1 xTd2 . This
proves the following lemma.
T T
Lemma 6.2.1 Let xa = xTa1 xTa2 and xd = xTd1 xTd2 be partitioned such that
xa1 and xd1 are the unknown variables matched by the set of constraints mx hx , then
there must exist explicite solutions for xa1 and xd1 . Let these solutions be denoted by,
The above Lemma states that a solution exists for a subset of the algebraic variables xa
and a subset of the state variables xd in the system, without introducing derivatives of
any known variables, i.e. variables in K. This solution can be used for elimination of
these variables. Moveover under the following assumption an output transformation can
be found, which has the property that the known and unknown variables in the output
expression are separated.
Assumption 6.2.1 It is assumed that the expression g2 in Lemma 6.2.1 exists and can
be rewritten to become,
T
ho (xd ) = go (y, u) , xd = xTd1 xTd2 . (6.6)
Assumption 6.2.1 states that the algebraic variables matched by the algebraic constraints
mx and hx can be eliminated using these. For linear systems, this is the same as assum-
ing that disturbances in the output equation are decoupled using a transformation of the
output, e.i. Qy = QCx + QEd d where QEd = 0.
Under Assumption 6.2.1 the following theorem can be used to derive a state space
description where all unknown algebraic variables are contained in the differential equa-
tions.
121
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Proof: Using Lemma 6.2.1 and Assumption 6.2.1 the set of constraints mx hx can be rewritten
to become,
Choose the second expression in (6.8) as the output equation, system (6.7) is obtained.
In some cases all the algebraic variables in the vector xa are match using the set of
constraints mx hx . When this is the case all the algebraic variables can be eliminated
using the output transformation given in Theorem 6.2.1. This is stated in the following
corollary.
Assumption 6.2.2 The state space description of system (6.7) is of the following form
dxd
dt = fo0 (xd , y, u) + Go (xd )xa2
(6.10)
go (y, u) = ho (xd ) ,
where Go (xd ) is a (n l2 ) matrix with full column rank for all xd Dxd . It is assumed
that n > l2 where l2 is the number of elements in xa2 .
122
Section 6.2: Realization
Later, in the proof of the main theorem in this subsection, it will be shown that, for the
over-constrained subsystems, Go has full rank always.
The elimination of the algebraic variables xa2 in (6.10) are dealt with in the follow-
ing Lemma.
Lemma 6.2.2 Eliminating algebraic variables xa2 from the state space description of
system (6.10) results in a new system on the form
D(xd ) dx 0
dt = fo (xd , y, u)
d
, (6.11)
Proof: Let the state space system (6.10) be scaled by an arbitrary (n n) matrix D(xd ), which
is full rank for all xd Dxd Rn . Using D on (6.10) the system becomes,
D(xd ) dx
dt
d
= D(xd )fo0 (xd , y, u) + D(xd )Go (xd )xa2 .
Design the scaling matrix D such that the following condition is fulfilled,
0
D(xd )Go (xd ) = . (6.12)
Il2 l2
The solution can easily be found using the Gauss-Jordan elimination method on the system,
0
Inn Go (xd ) D(xd ) .
Il2 l2
In this expression the first (n l2 ) rows are independent of xa2 . This means that xa2 has been
eliminated in this part of the system. Using the first (n l2 ) rows in (6.13) the final expression is
obtained,
D(xd ) dx
dt
d
= D(xd )fo0 (xd , y, u) = fo0 (xd , y, u)
The matrix D(xd ) is formed by the first (n k2 ) rows of the matrix D(xd ).
The following additional assumption, that is required to prove the main theorem, is
introduced.
123
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Remark 6.2.1 A linear version of (6.10) is given by the following state space model
dxd
dt = Axd + Bu + Gxa2
(6.15)
y = Cxd .
Theorem 6.2.2 (Realization) By a state transformation, system (6.10) can, under as-
sumption 6.2.3, be transformed into a new system of the following form
dz
dt= fz (z, y, u) = (xd )fo0 (xd , y, u)|xd =(z,y,u)
(6.16)
go1 (y, u) = hz (z, y, u) = ho1 (xd )|xd =(z,y,u) .
The inverse state transformation xd = (z, y, u) is obtained as the local solution to
go2 (y, u) ho2 (xd )
= , (6.17)
z (xd )
with output map organized as
go1 (y, u) ho1 (xd )
go (y, u) = ho (xd ) = .
go2 (y, u) ho2 (xd )
The transformation z = (xd ) is given by the partial differential equation (p.d.e.),
(xd )D(xd ) = , (6.18)
xd
where D is given by Lemma 6.2.2 and (xd ) is a l2 l2 scaling matrix with full rank
for all xd Dxd Rn .
Proof: The proof is done in two steps. In the first step it is shown that all xa2 can be eliminated
in (6.10). In the second step it is shown that under Assumption 6.2.3 there exists a transformation
transforming system (6.10) into (6.16).
Step 1: First it is shown that (6.10) is on a form such that Lemma 6.2.2 can be used in the
elimination of the unknown algebraic variables xa2 . To show this, it must be proven that Go (xd )
has full column rank and that the number of rows exceed the number of columns.
System (6.7) is by definition a over-constrained subsystem with the set of constraints Cmin =
do fo ho and the set of unknown variables Xmin = xd xd xa2 . Using Definition 6.1.7
the following property must hold for (6.7)
|do | + |fo | + |ho | = |xd | + |xd | + |xa2 | + k ,
124
Section 6.2: Realization
where | | is the number of elements in the given vector and do is the set of differential constraints
(arranged in a vector) as described in Section 6.1. By definition |do | = |xd |. Moreover, some
of the elements in xd must be matched using ho . Let these elements be denoted xd1 . Since all
constraints in ho are matched we have |xd1 | = |ho |, and we get
D(xd ) dx
dt
d
= fo0 (xd , y, u)
(6.20)
g(y, u) = ho (xd ) .
A method for calculating D is given in the proof of Lemma 6.2.2. Multiplying the differential
equation (6.20) with the scaling matrix , we obtain an expression for dz
dt
dz dxd
= (xd )D(xd ) = (xd )fo0 (xd , y, u) . (6.21)
dt dt
From this equation it is seen that is given by the p.d.e.
= (xd )D(xd ).
xd
Step 2.: Next it is shown that there exists an expression on the form
go2 (y, u) ho2 (xd )
= , (6.22)
z (xd )
where go2 (y, u) = ho2 (xd ) contains a subpart of the rows in go (y, u) = ho (xd ). This expres-
sion must be solvable for all xd in the system, i.e. a local solution must exists. For this to be true
the inverse function theorem states that the Jacobian determinant of (6.22) must be different from
zero, e.i.
ho2 (xd ) !
ho2 (xd ) xd
= (6.23)
xd (xd ) (xd )D(xd )
xd = (z, y, u).
125
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Remark 6.2.2 The question arise about the existence of the solution z = (x) that
satisfies (6.18). In R2 the problem can be formilized by
= (x)D(x) x
x 1 x2 = (x) d1 (x) d2 (x)
2 2 d1 d2
x1 x2 x1 x2 = x2 x1 =0
This expression can be solved for (in the R2 case) using the intergrating factor ap-
proach (Nagle and Saff, 1996, chap. 2). When is known can be found by simple
intergration (this approach is used in the example in Section 6.2.4).
The expression described above, can in R3 be described by
where i denotes the ith row in (6.18). Using the operator the following expression
can be obtained
Remark 6.2.3 In general the existence of a solution to the p.d.e. (6.18) can be treated
using Frobenius theorem (Isidori, 1995). A discussion on this approach is found in (Frisk
and slund, 2003), which treats the solution to problems similar to (6.18).
1. First use Theorem 6.2.1 to eliminate a subset of the unknown algebraic variables
xa . If all the algebraic variables can be eliminated using this approach Corollary
6.2.1 states the final solution. If this is not the case go to step 2.
126
Section 6.2: Realization
2. Check if Assumptions 6.2.2 and 6.2.3 holds for the system. If it holds go to step 3,
else it is not possible to transform the system into a state space description using
the approach given here.
3. Use Theorem 6.2.2 to obtain the state transformation , and derive the trans-
formed system.
4. Finally, to design a residual observer, find a stabilizing feedback for the trans-
formed system. This is not treated here.
In this section ideas for developing residual observers using over-constrainted subsys-
tems found using SA are presented. Even though deriving residual observers were the
aim of the section, it is possible to use the same ideas to identify subsystems which can
be used in the development of reduced order observers. The following theorem can be
used to formulate a necessary condition for this to be possible.
is observable for all x Dx , then there exists an over-constrained match of the unknown
variables X = x x using the system constraints C = h f d.
Proof: A necessary but not sufficent condition for system (6.24) to be observable is,
f
(x)
rank h x =n x Dx .
x
(x)
This implies that, when (6.24) is observable then all xi i {1, , n} are containted in either
f or h. This implies that there exist a match of all xi . When this match exist xi can be match
using the differential constraints d. The length of both f and d is n, and the number of unknown
variables X is 2n. Therefore, there exist |h| more constraints than unknown variables, i.e. the
system is over-constraint. This completes the proof.
Remark 6.2.4 Theorem 6.2.3 states that if system (6.24) is observable, then there exist
and over-constrained match on the system. Reversing this argument, it can be said
that if there does not exist an over-constraints subsystem, then there does not exist an
observable state space description of the system.
Remark 6.2.5 The arguments presented in this section does not garantee the existence
of a residual observer or a reduced order observer, as the stability of an observer imple-
mentation is not considered here.
127
Chapter 6: A New Approach for FDI in Centrifugal Pumps
c1 : = v
c2 : v = 2 1 12 + 2 u1 + d
c3 : =
u2 f
c4 : = 2v
+ 2 + 2
(6.25)
c5 : y1 =
c6 : y2 =
c7 : y3 =
and four derivative constraints of the form x = dx dt . In this model (, ) denotes the
position of the satellite in polar coordinates on the plane, v is the radial velocity, is
the angular velocity and u1 , u2 are the radial and tangential thrust, respectively. f is the
fault signal and d represents a disturbance signal. The parameters 1 and 2 are supposed
to be known, constant and different from zero. The constraints c1 to c4 describe the
dynamics of the satellite and the constraints c5 to c7 describe the measurement system
on the satellite.
The structural graph of the satellite system (6.25) is shown in Table 6.3. From this
Table 6.3: The structure table of the satellite system. x means uni-directional relations
and 1 means bi-directional relations, where uni-directional means that the given variable
is not calculable from the relation, see definitions in (Izadi-Zamanabadi, 2001). The
symbols show a matching for a part of the system.
Known Unknown Faults
y1 y2 y3 u1 u2 d v v f
c2 1 1 1 1 1
d4 1 x
d2 1 x
c3 1 1
c6 1 1
d1 1 x
c1 1
c4 1 1 1 1 1
d3 x
c5 1
c7 1
table it is seen that a match exists, which includes the constraint affected by the fault f .
Therefore the subsystem formed by this match can be used for fault detection. Moreover
the subsystem does not contain the disturbance d, and is therefore robust with respect to
128
Section 6.2: Realization
C = {d1 , d3 , c1 , c4 , c5 , c7 }.
Considering the theory developed in Section 6.2, the set of equations describing the iden-
tified subsystem must be described as in (6.4). The state, algebraic and output vectors
are therefore given by
T T
xd = xa = v y = y1 y3 ,
Form these expressions it is seen that the the vector field fx must be used for the elimina-
tion of xa . fx can be put on the form fo0 (xd , u) + Go (xd )xa , meaning that Assumption
6.2.2 is fulfilled. Duing this the expression becomes
0 1
= + v. (6.26)
2 u2 2
Hence, Lemma 6.2.2 can be used to obtain the matrix D as it is shown in the following
by choosing
d11 (xd ) d12 (xd )
D(xd ) = , (6.27)
d21 (xd ) d22 (xd )
we compute D(xd ) dx 0
dt = D(xd )fo (xd , u) + D(xd )Go (xd )xa and obtain
d
dxd d12 (xd )2 u2 d11 (xd ) 2d12 (xd )
D(xd ) = + . (6.28)
dt d22 (xd )2 u2 d21 (xd ) 2d22 (xd )
u2
2
2 =
0
y= .
129
Chapter 6: A New Approach for FDI in Centrifugal Pumps
= 2 .
When exists and Assumption 6.2.3 is fulfilled then the state transformation is,
according to Theorem 6.2.2, given by the explicit solution to,
y1
= ,
z 2
where y1 = is one of the output constraints, see (6.25). The solution to this equation
becomes,
z/(y12 )
== .
y1
Using this transformation the state space representation of the identified subsystem be-
comes,
z = 2 y1 u2
S:
y3 = z/(y12 ) .
where k and q are design constants. The observer O is exactly the same observer as
obtained in (De Persis and Isidori, 2001) using the geometric approach.
130
Section 6.3: System Model
Application
vsabc Centrifugal Qp
pump Hp
Figure 6.2: The centrifugal pump placed in a hydraulic application. The inputs to the
centrifugal pump are the supply voltage vsabc and the volume flow Qp and the output is
the pressure difference Hp . The inputs to the application are a set of unknown inputs d
and the pressure difference of the pump Hp , and the output is the volume flow Qp .
vsabc and the volume flow Qp , and the output is the pressure difference Hp delivered by
the pump. This input/output structure is depicted in Fig. 6.2 where it is connected to a
general hydraulic application. The inputs to this application are a set of unknown inputs
d and the pressure difference Hp of the pump, and the output is the volume flow Qp .
The model of the system, depiced in Fig. 6.2, is formalized in the following subsections.
where xQ is a vector containing the states of the hydraulic application and the vector d
represents the unknown inputs to this application. Hp and Qp are the pressure output
and the volume flow of the centrifugal pump respectively.
The model of the centrifugal pump, presented in Chapter 3, is connected in a feed-
back loop with the application model, as it is depiced in Fig. 6.2. This results in the
following model,
dxd
dt = fx (xd , u, d)
y= hx (xd ) ,
131
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Finally, the unknown input vector d is the same vector as the unknown input vector
prestented in (6.29). The vector field fx of the system is described by the following set
of equations,
R +R0 R0 L0
sL0 r isd + L0r imd zp r Lm0 imq + L10 vsd
s
R +R0
s
L0
s
R0
s
sL0 r isq + zp r Lm0 imd + L0r imq + L10 vsq
s
Rr0
Rr0
s s s
i i + z i
fx (xd , u, d) = L0m sd L0m md p r mq ,
Rr0 Rr0
i
L0m sq z i
p r md i
L0m mq
1 3
zp LM (imd isq imq isd ) r fT (h(xQ ), r )
0 B
J 2 J
fA (xQ , H, d)
This model represents the behaviour of the system when no fault has occured. In the
following the effect of faults on the centrifugal pump is described.
132
Section 6.4: Structural Analysis
The mentioned faults all affect the hydraulic part of the pump. The performance of
the hydraulic part of the pump is modelled by fH and fT . These functions describe
the pressure and the torque produced by the pump respectively. Moreover the flow
measurement is part of the hydraulic description of the pump. Introducing the faults
described above, the description of the hydraulics of the pump becomes,
As stated in Section 6.1 the structural analysis is the study of the system properties,
which are independent of the actual values of the parameters. Only links between the set
of variables X and the set of constraints C are represented in the analysis. In this work
the use of SA is two folded, as it is both used for identifying two cascade connected
subsystems, and to identify subsystems which can be used for residual generations.
The system is splitted into two cascade connected subsystems to facilitate the deriva-
tion of residual generators. This is possible as the faults considered in this work only
affect the second subsystem, meaning that only this subsystem must be considered when
deriving residual generators.
As the SA is working on the two sets C, Z these must be defined for the pump
application before the results of the SA can be obtained. This is done in the Section
6.4.1. After that the splitting of the system is considered in Section 6.4.2. Finally this
section ends with identifying subsystem for residual generation in Section 6.4.3.
The constraints C of the hydraulic system are identified from the model presented in
Section 6.3. These constraints are given by
133
Chapter 6: A New Approach for FDI in Centrifugal Pumps
where the constraints c7 and c8 are included to make the constraints c1 to c4 independent.
Hereby Assumption 6.1.3 is fulfilled globally (Blanke et al., 2003).
In (6.31) the constraints c1 to c4 and c7 to c8 describe the electrical part of the
induction motor, c5 describes the mechanical dynamics of the pump, and c6 describes
the dynamics of the hydraulic applications. c9 , c10 and c11 describe the torque generated
by the induction motor, the pressure difference generated by the centrifugal pump, and
the load torque of the centrifugal pump respectively. c12 describes the volume flow
through the pump and finally c13 to c16 describe the measurements on the system.
Two extra constraints can be deduced by differentating c7 and c8 with respect to
time. Doing this the following additional constraints are obtained,
Beside the constrains presented above, there is a differential constraint for each variable
xd , meaning that a constraint on the form,
dxdi
di : = xdi , (6.33)
dt
exists for each element in xd . In this expression i denotes the ith element in xd .
From the constraints presented in (6.31) and (6.32) the set of variables are identified.
These are given by,
Z = xd xd xa u y
134
Section 6.4: Structural Analysis
Ce is defined as the first subsystem. The remaining relations are defined as the second
subsystem, meaning that the constrains of this subsystem are given by,
The connecting variables between the two subsystems are in this work defined as the
estimates of r and Te , and will in the following be denoted r and Te respectively.
It is also seen from the column at the right hand side of Table 6.4 that the faults
treated in this chapter are not affecting the relations in the first subsystem. Therefore, it
is only necessary to consider the second subsystem when designing residual generators
for FDI on the system. Hence, the fault detection algorithm can be divided into two parts
as shown in Fig. 6.3.
Using the relations describing the first part, an adaptive observer is designed. This
observer observes the variables r and Te connecting the two parts. The design of this
observer is considered in Section 6.5. But first SA is used to identify minimal over-
constrained subsystems (see Definition 6.1.11) in the system formed by the constraints
Cm . The obtained results are presented in the following subsection,
135
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Table 6.4: The structure table of the system. x implies that the given variable is not cal-
culable using the constraint but does appear in the constraint (Izadi-Zamanabadi, 2001).
The symbols show a matching for the first part of the system.
isd isq Kf Kl B fc fd
1
1
1
1
Faults
1
1
x
1
x
y1 y2 y3 y4 vsd vsq uQ xQ xQ Tp Hp Qp Te r r imd imq imd imq A1 A2 A1 A2 isd isq
1
1
1
x
1
x
1
1 1
1
x
1 1
1
Unknown
1 1 1 1
1 1 1
1
1 1 1
x 1
1
1
1
1
1
1
1 1
1
1
1
1
1
1 1 1
1 x
1
1
1
Known
1
1
1
1
dd1
dd2
cd8
cd7
c10
c13
c14
c14
c15
d6
d5
d3
d4
d1
d2
c6
c5
c9
c8
c9
c3
c4
c7
c8
c1
c2
136
Section 6.4: Structural Analysis
Observer
Te FDI
vabc
First
iabc subsystem Second
r subsystem
r
H
Q
Figure 6.3: The division of the fault detection algorithm. In the first subpart the model
of the electrical part of the motor is used, and in the second subpart the models of the
mechanical and hydraulic parts of the system are used.
The SA is in this section used for identification of four minimal over-constrained sub-
systems. Each of these subsystem can be utilized for detecting a subset of the faults
in the centrifugal pump. The second subsystem is described by the set of constraints
Cm in (6.35). Beside these constraints two extra constraints are defined to describe the
connecting variables r and Te . These constraints are given by,
c15 : r = r
c16 : Te = Te .
In Table 6.5 the graph of the second subsystem is shown. Here the two extra constraints
are added.
Table 6.5: The structural model of the second subsystem Cm obtained in Section 6.4.2.
Known Unknown Faults
y3 y4 r Te uQ xQ xQ r Tp r Hp Qp Te Kf Kl B fc fd
c6 1 1 1 1
d6 1 x
c10 1 1
d5 1 x
c5 1 1 1 1
c9 1 1 1 1 1 1
c8 1 1 1 1 1 1
c15 1 1
c13 1 1
c14 1 1 1
c16 1 1
Using the definitions and procedures described in (Blanke et al., 2003; Izadi-
Zamanabadi, 2001), four minimal over-constrained subsystems are identified. The set
137
Chapter 6: A New Approach for FDI in Centrifugal Pumps
138
Section 6.5: Observer for the Motor Part
which fulfills the demand of existence given in Remark 6.2.4. The new set C0e is formed
by the following constraints,
In this set the unknown variables xd xa and the known variables u y are given by,
T T
xd = isd isq imd A1 A2 xa = imq imq r r
T T
u = vsd vsq y = y1 y2 .
Comparing the set (6.38) with the general system given by (6.4) the vector field fx , the
algebraic maps mx and the output maps hx are constructed using the set of constraints
{c1 , c2 , c3 , cd7 , cd8 }, {c4 , c7 , c8 }, and {c13 , c14 } respectively. According to Theorem
6.2.1 and the match shown in Table 6.4, it is possible to eliminate a subset of the alge-
braic variables xa in the system using the vector function mx . From the match presented
139
Chapter 6: A New Approach for FDI in Centrifugal Pumps
in Table 6.4 it is seen that the two algebraic variables imq and imq should be eliminated
using mx . By eliminating these the following system is obtained,
dxd
dt= fo0 (xd , xa2 , u)
(6.39)
go (y) = ho (xd , xa2 ) ,
T
where xa2 = r r and,
Rs +Rr0 1 1
L0s isd + L0s A1 + L0s vsd
Rs +Rr0
L0 isq + L10 A2 + L10 vsq
s s
R0
s
fo =
0 L10 A1 + L0r isd
0 m m
Rr 0 L L0 2
L0 (Rr isd A1 ) + zp r Rm0 A2 + zp r Rm0 imd + zp r (Rr0 isq A2 )
m r r
R0 Rr0 2 0
L0r A2 + L0 sq i z L
p r m mdi zp r (Rr0 isd A1 )
m m
L0 2 L0
0 Rr0 + zp2 r2 Rm0 imd + zp r Rm0 A2 A1
r r
go = y1 ho = isd .
y2 isq
The next step in the derivation of a state space description is to use the approach de-
scribed in Theorem 6.2.2 to eliminate the algebraic variables xa2 in (6.39). Unfortu-
nately, doing this the expression becomes huge, making it very difficult to find the state
transformation , see Theorem 6.2.2. This makes the obtained expression useless.
However, by eliminating the variables in xa2 in two steps an useful expression
is obtained. To see this first r is eliminated in (6.39), and then the state trans-
formation T : x, r xd is used to transform the obtained system. Here x =
T
isd isq imd imq and T is given by
isd
isq
T(x, r ) = imd ,
Rr0 imd + zp r L0m imq
Rr0 imq zp r L0m imd
Doing this the traditional description of the electrical part of an induction motor is ob-
tained, i.e. the transformed model has the following form
dx
dt = fx (x, r , u)
(6.40)
y = hx (x) ,
140
Section 6.5: Observer for the Motor Part
where
R +R0 R0 L0 1
sL0 r isd + L0r imd zp r Lm0 imq + L0s vsd
R +R0
s s s
s 0 r i + z L0m0 i + R0r0 i + 1
sq p r md Ls mq L0s vsq i
fx (x, r , u) = Ls
Rr0
Ls
Rr0 hx (x) = sd .
isq
L0m isd L0m imd + zp r imq
Rr0 Rr0
L0 isq zp r imd L0 imq
m m
As the second step the approach presented in Theorem 6.2.2 is used ones again. This
time to eliminate r in (6.40). Hereby the following state space description is obtained,
dz
= fz (z, u0 )
dt (6.41)
y = hz (z, u0 ) ,
T
where z Dz R3 and u0 = vsd vsq y2 . The vector field fz and the nonlinear
map hz in this expression are given by,
fz (z, u0 ) =
Rs 2L0m z3 z22 +2z2 L0s y2 L0s 2 y22
R
L0 z1 +
s
+ vsd
s L0s
Rs y2 + vsq
0
0 z z 2 +2z L0 y L0 2 y 2 z
R0 L0 +R0 L0 Rr0 2 R 0 R 2L 3 2 2 1
2 r Ls0 L0 r m z3 + L0 L0 z2 L0 z2 y2 +
r r m 2
L0 L0
s s 2
s m m s m m s
2L0m z3 z22 +2z2 L0s y2 L0s 2 y22
hz (z, u0 ) = 1
L0s z1 L0s
.
The system (6.41) is a state space realization of the set of constraints C0e , where all
algebraic variables are eliminated. The speed of the motor can be calculated from the
states of (6.41) using the following expression,
R0r R 0 L0
L0m
z2 +vsq L0s y2 Rs + Lr0 s +Rr0 y2
r = m
,
zp 2L0m z3 z22 +2z2 L0s y2 L0s 2 y22
from which it is seen that r is a function of y2 , which in general is not known. Even
though this expression is not usable for estimating the speed r , the following remark
on this system should be considered.
Remark 6.5.1 If an observer can be found based on system (6.41), then this observer
is a residual observer for the induction motor, which is independent of the speed r .
In this section it is argued that it is not possible to estimate the speed of the motor without
knowing the derivative of at least one of the measurements. However, it is also shown
that the subsystem identified using SA corresponds to the electrical part of an induction
motor, see (6.40). Therefore, all the methods, described in the literature, for speed
and torque estimation based on the electrical model can be used. In the following the
adaptive observer, developed in Section 5.2.1, is utilized for estimating the connecting
variables, i.e. motor speed and torque.
141
Chapter 6: A New Approach for FDI in Centrifugal Pumps
where the state vector x, the input vector u and the output vector y are given by
T
x = isd isq imd imq
T T
u = vsd vsq y = isd isq ,
1
L0s 0
0 1
1 0 0 0
B=
0
L0s C= .
0 0 1 0 0
0 0
If the speed is assumed constant, this system is a linear system with one unknown but
constant parameter. Using the transformation x = Tz given by
1 0 0 0
0 1 0 0
T = L0s 0 1 0
L0 m
L0
0 L0s 0 1
m
the system (6.42) is tranformed to the adaptive observer form defined in Definition 5.2.1
in Section 5.2.1, where A(u, ) = A() and = r . This means that an adaptive
142
Section 6.6: Observer Based Fault Detection and Isolation
observer is given by Proposition 5.2.1. Using this proposition the adaptive observer for
the induction motor becomes
dz
dt = (A00 + r A0r )z + B0 u + K(y C0 z)
dr
Oe : = (y C0 z)T A00r z (6.43)
dt
x = Tz ,
where z R4 contains the observer states, and r is the estimated speed. The matrices
in the observer are
A00 = T1 A0 T A0r = T1 Ar T
B0 = T1 B C0 = CT ,
and A00r is a matrix composed by the two first rows of A0r . K and are design
constants. K is chosen such that the LMI (5.27), described in Section 5.2.2, is feasible,
and is chosen such that the convegence speed of r is suitable.
The connecting variables between the two subsystems are the speed and torque. Of
these only the speed is directly available from the observer (6.43). However, using c9
the torque can be calculated whenever the states of the motor are known. When the
estimates of the states are used c9 becomes,
c9 : Te = 1.5zp L0m imd isq imq isd ,
where isd , isq , imd and imq are estimates of states in the original system presented in
(6.42).
This observer is, as mentioned in the beginning of this subsection, developed under
the assumption that the speed is constant i.e. d dt = 0. This assumption is not correct
r
during transient phases, therefore it is expected that problems can arise when transients
occur in the speed.
143
Chapter 6: A New Approach for FDI in Centrifugal Pumps
It is also possible to obtain ARRs from the sets Cm1 , Cm3 and Cm4 , but because a
differential constraint is used in each of these sets, it is necessary to use derivatives of
the output in these cases. To avoid this, three residual observers are developed in the
following. The constraints contained in the sets Cm1 , Cm3 and Cm4 are given by (6.36),
and are reproduced below for convenience,
d5 : r = d
dt
r
c5 : J r = Te Br Tp
c8 : Hp = ah2 Q2p + ah1 Qp r + ah0 r2
c9 : Tp = at2 Q2p + at1 Qp r + at0 r2
c13 : y3 = Hp
c14 : y4 = Qp
c15 : r = r
c16 : Te = Te .
Comparing the sets Cm1 , Cm3 and Cm4 with the structure of (6.4), the vector field fx
is formed by the constraint c5 , the map mx is formed by a subset of {c8 , c9 }, and the
map hx is formed by a subset of {c13 , c14 , c15 , c16 }. Using Corollary 6.2.1 on these sets
it is shown that the sets Cm1 , Cm3 and Cm4 all can be transformed to systems with the
following structure,
dxd
dt = axd + fx (xd , v1 , u) + e1 (xd , v1 , u, f )
(6.45)
g(v1 , v2 , u) = h(xd ) + e2 (xd , v1 , v2 , u, f ) ,
Assumption 6.6.1 It is assumed that in the case where no faults have occurred, i.e.
f = 0, the output map h in (6.45) can be solved for xd locally. Let this solution be given
by,
xd = g(v1 , v2 , u) . (6.46)
144
Section 6.6: Observer Based Fault Detection and Isolation
Using the above assumption the following proposition describes a residual observer
for the system described by (6.45).
Proposition 6.6.1 Under Assumption 6.6.1 the following observer is a residual ob-
server for systems described by (6.45),
xd
= axd + fx (g(v1 , v2 , u), v1 , u) + k(g(v1 , v2 , u) xd )
dt . (6.47)
r = q(g(v1 , v2 , u) xd )
The residual observer is asymptotical stable if ak < 0. The fault input to this observer
is given by,
where (6.46) is used in the observer expression (6.47), meaning that g(v1 , v2 , u) = xd . Equation
(6.48) shows that the error dynamic of the observer is asymptotical stable if a k < 0.
The expression of the derived fault signal ff is obtained in the following by introducing the
fault signals in the error equation of the observer. First, an expression of the fault when mapped
through g is obtained. To this end define the function gf as,
From this expression it is seen that the function gf must be used to obtain a map from the mea-
surements v1 and v2 to the state xd if a fault f has occured, i.e.
From this expression the following nonlinear expression of the fault can be identified,
145
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Remark 6.6.1 The derived fault ff is strongly detectable using this observer. This is
not the case for the faults in f , as the nonlinear expression of ff can equal zero even
though one of the entries in f is different from zero.
Remark 6.6.2 The observer described by Proposition 6.6.1 is designed under the as-
sumption that a perfect model exists, and that the measurements are not affected by
noise. This is, of course, not fulfilled in real life applications. To overcome this the gain
k of the observer is chosen such that errors due to small model mismatchs and noise will
be suppressed.
Using Corollary 6.2.1 and Proposition 6.6.1 on the three over-constrained subsystems
Cm1 , Cm3 , and Cm4 , the following three residual observers are obtained,
dx
J dtd = B xd fT (y4 , r ) + Te + k1 (r xd )
Om1 : (6.50)
r1 = q1 (r xd )
J dx
dt = B xd fT (g3 (r , y3 ), r ) + Te + k3 (r xd )
d
Om3 : (6.51)
r3 = q3 (r xd )
J dx
dt = B xd fT (y4 , g4 (y3 , y4 )) + Te + k4 (g4 (y3 , y4 ) xd )
d
Om4 : (6.52)
r4 = q4 (g4 (y3 , y4 ) xd )
where ki is designed according to Proposition 6.6.1 and qi is chosen such that the resid-
uals have a reasonable value in the case of faults. The function fT is given by
fT (Qp , r ) = at2 Q2p + at1 Qp r + at0 r2 ,
and the functions g3 and g4 are derived from the output map h in (6.45), and are given
by,
p
ah1 r + a2h1 r2 4ah2 (y3 ah0 r2 )
g3 (r , y3 ) =
2ah2
p
ah1 y4 + ah1 y42 + 4ah0 (y3 + ah2 y42 )
2
g4 (y3 , y4 ) = .
2ah0
These expressions are valid for r , y4 R+ when the parameters of the pump used in
the test are considered. Therefore the expressions are valid in the state space r , Q
R+ , which is exactly the state space in which the model is valid, see Section 3.5.
146
Section 6.7: Test Results
vsabc Te
Motor r1
Observer ARR and
isabc r2
r Residual
Observers r3
Hp r4
Qp
Figure 6.4: The final detection algorithm. The first part is the motor observer described
in Section 6.5, and the second part is composed of the ARR and residual observers
described in Section 6.6.
estimating the connection variables r and Te , based on the supply voltage vsabc and
the motor current isabc . These connection variables are then, together with the flow and
pressure measurement Qp and Hp , used as input to the residual generators.
This detection algorithm is in the following tested on a Grundfos 1.5 (KW ) CR5-10
pump. This pump is placed in a tank system as depicted in Fig. 6.5. In this tank system
tank
Hp
Vl
V1 V2
pump V3
Vc
pipe Qp
Te
shaft
M
motor
Figure 6.5: Sketch of the test setup. The measurements are the electrical quantitatives,
the differential pressure Hp delivered by the pump and the volume flow through the
pump Qp .
the valve V1 is used to model disturbances in the system. Clogging inside the pump is
147
Chapter 6: A New Approach for FDI in Centrifugal Pumps
modelled by the valve Vc and dry running is modelled by closing V2 and opening V3 .
Rub impact is modelled by adding an extra force to the shaft and cavitation is modelled
by closing valve V2 gradually. Leakage flow is modelled by opening Vl .
Test results have shown that the sensitivity to the faults fc and fd of the observer
Om4 is very low. Infact it is so low that changes due to the faults are smaller than
changes due to noise and parameter variations. Moreover in Section 6.4.3 it is shown
that the obtainable fault information is included in the residuals r1 , r2 , and r3 . Therefore
only these residuals are considered in the test presented in this section.
Since the tests are performed on a real system, noise is expected on the residuals. To
overcome this problem a CUSUM algorithm (Basseville and Nikiforov, 1998) is used to
detect changes in the mean of the residuals and thereby detect the faults. The CUSUM
algorithm is shortly described in Appendix B.1. In the following, outputs of the CUSUM
algorithms are denoted D1 to D3 , where D1 is the decision signal of r1 and so forth.
All test results are shown in Fig. 6.6, Fig. 6.7 and Fig. 6.8. First robustness
with respect to the operating point is tested. In this test both the position of the valve
V1 and the speed of the pump are changed during operation. During the test the valve
is changed in three steps from medium to maximum opened. The speed of the pump
is changed between 2380 and 2910 (rpm) each 2 (sec) during the test. The result of
this test is shown in Fig. 6.6(a), where r1 to r3 are shown in the top figure and the
decision signals D1 to D3 in the bottom figure. The test shows that the three residual
generatores are robust with respect to the tested operating points, but also that there are
some dependency to the operating point, see top figure of Fig. 6.6(a), This is partly due
to problems with the flow sensor at zero flow and partly due to dependency between the
parameters and the operating point.
Figures 6.6(b) to 6.8(b) show test results concerning isolability of the five faults
considered in this chapter. All these tests are performed with V1 half opened and an
angular speed of approximately 2650 (rpm). Comparing D1 , D2 , and D3 in the five
figures it is seen that the faults are distinguishable except for cavitation and dry running
shown in Fig.6.8(a) and Fig. 6.8(b) respectively. This was expected as the structural
analysis in Section 6.4.3 already had predicted this.
6.8 Discussion
The first topic of the work presented in this chapter is realization of over-constrained
subsystems identified using Structural Analysis (SA). It is well known that there is a
straightforward connection between Analytical Redundant Relations (ARRs) and mini-
mal over-constrained subsystem. Unfortunately, the obtained ARRs are in general func-
tions of the derivatives of the measurements, which are difficult to calculate when the
measurements are corrupted by noise.
To overcome this problem a new method for rewriting a subsystem identified using
SA to a state space description is developed. The obtained state space description has the
property, that the only unknown variables are the states of the system. The state space
148
Section 6.8: Discussion
Robustness
1.5 r1
r2
1 r3
0.5
residuals 0
0.5
1
1.5 Valve opening step 1 Valve opening step 2
0 2 4 6 8 10
time [sec]
3
D1
2.5 D2
D3
2
Decisions
1.5
0.5
0
0 2 4 6 8 10
time [sec]
Clogging
1.5 r1
1 r2
r3
0.5
residuals
0.5
1.5
0 2 4 6 8 10 12 14 16 18
time [sec]
3
D1
2.5 D2
D3
2
Decisions
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18
time [sec]
Figure 6.6: Test results from test of the developed algorithms on the test setup. The
top figures shows the obtained residuals and the bottom figures shows decision signals
obtained from CUSUM algorithms.
149
Chapter 6: A New Approach for FDI in Centrifugal Pumps
Leakage Flow
r1
2 r
2
r3
1
residuals
0
0 2 4 6 8 10
time [sec]
3
D1
2.5 D2
D3
2
Decisions
1.5
0.5
0
0 2 4 6 8 10
time [sec]
Rub Impact
2 r1
r
2
1 r3
residuals
2
0 2 4 6 8 10 12 14 16
time [sec]
3
D1
2.5 D
2
D3
2
Decisions
1.5
0.5
0
0 2 4 6 8 10 12 14 16
time [sec]
Figure 6.7: Test results from test of the developed algorithms on the test setup. The
top figures shows the obtained residuals and the bottom figures shows decision signals
obtained from CUSUM algorithms.
150
Section 6.8: Discussion
Cavitation
r1
r
5 2
r3
residuals 0
5 End of cavitation
0 2 4 6 8 10 12 14 16 18
time [sec]
3
D1
2.5 D2
D3
2
Decisions
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18
time [sec]
Dry Running
15
r1
10 r2
r3
5
residuals
10
15
0 2 4 6 8 10 12 14 16 18
time [sec]
3
D1
2.5 D2
D3
2
Decisions
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18
time [sec]
Figure 6.8: Test results from test of the developed algorithms on the test setup. The
top figures shows the obtained residuals and the bottom figures shows decision signals
obtained from CUSUM algorithms.
151
Chapter 6: A New Approach for FDI in Centrifugal Pumps
description can then be used for derivation of residual observers. In the design of these
observers decoupling of unknown inputs has not to be considered, as the only unknown
variables are the states of the system. The method is tested on a satellite model, showing
that a residual observer can be obtained using this approach. Moreover the method is
used to develop residual observer for the centrifugal pump application, which is the
second topic of the chapter.
The second topic of the chapter is fault detection and isolation in a centrifugal pump
placed in an arbitrary hydraulic application. An algorithm, which is capable of detection
and isolation of five faults in a centrifugal pump, is developed. The proposed alogrithm
is independent of the application in which the pump is placed. This makes the algorithm
robust and usable in a wide range of applications, such as submersible application, waste
water application, and heating application.
Tests have shown that it is possible to distinguish between four of the five faults un-
der consideration, using three chosen residuals. But it is also shown that the algorithm is
sensitive to the operating point. This is partly due to dependency between the operating
point and the parameters in the model and partly due to flow sensor problems at zero
flow. Even though the algorithm has a small inherent dependency of the operating point,
it still performs considerably better than algorithms built on a linearized model, when
the operating point is changed.
152
Chapter 7
In this chapter Structural Analysis (SA) is utilized to derive Analytical Redundant Re-
lations (ARRs), which can be used for fault detection. The concept of SA is described
in Section 6.1 in the previous chapter. The obtained ARRs are based on a steady state
model of the pump. Here a steady state model denotes a model describing the pump
under constant speed, pressure and flow conditions. The developed detection algorithm
is only using electrical measurements on the motor, and the pressure and flow measure-
ments on the centrifugal pump. The electrical measurements are in this case Root Mean
Square (RMS) measurements of the voltage and current, the voltage frequency, and the
electrical angle between the voltage and current. These measurements become constant
when the system is running under steady state conditions, and can therefore be sampled
at any given sample rate, provided the steady state conditions. Therefore, by using only
these measurements the microprocessor load of the derived algorithm can be chosen
freely. Moreover, the steady state measurements of the electrical signals are sometimes
made available by modern motor protection units. All in all this makes the algorithm
suitable for implementation in cost sensitive products.
The obtained algorithm is robust with respect to parameter variations and the oper-
ating point of the pump respectively, making it usable in real life applications. In the
development of the FDI algorithm, the approach, presented in Section 6.4, of dividing
the system into two cascade-connected subsystems is used. This is done to avoid two
large residual expressions. The first of these subsystems consists of the electrical part
of the induction motor driving the pump, and the second subsystem consists of the me-
chanical and hydraulic part of the pump. Theoretical considerations regarding to Struc-
tural Analysis (SA) are, among others, found in (Blanke et al., 2003; Izadi-Zamanabadi,
2001).
The faults, which only affect the second subsystem, are detected using Analytical
153
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
Redundancy Relations (ARRs). These relations are obtained through the utilization of
structural analysis (Blanke et al., 2003) and the Groebner basis algorithm (Cox et al.,
1997). Using this approach the obtained ARRs are polynomial. This is utilized in
the development of a detection algorithm, which is robust with respect to parameter
variations. Theoretical results on using the Groebner basis are given in (Staroswiecki
and Comtet-Varga, 2001). Moreover, an overview of model-based methods using ARRs
is given in (Staroswiecki, 2000).
This chapter starts by presenting the steady state model of a centrifugal pump placed
in an arbitrarily hydraulic application. This is done in Section 7.1. Structural analysis on
the derived steady state model is considered in Section 7.2, which includes; the division
of the system into two suitable subsystems, derivation of an expression for calculating
the connecting variables, and ARR expressions for fault detection. In Section 7.3 a
method for robust fault detection using ARRs is considered. Section 7.4 presents some
test results obtained on the industrial test bench, which also was used to obtain the test
results presented in Chapter 6. Finally concluding remarks end the chapter.
The variables vsdq , isdq , and imdq are in general unknown. However, in the Y-connected
case the measurable electrical quantities at the terminals of the motor itdq and vtdq
equals isdq and vsdq respectively. In the -connected case these currents and voltages
are given by the transformations itdq = Ci isdq and vsdq = Bv vtdq respectively, see
Section 3.2.3.
154
Section 7.1: Steady State Model of the System
To obtain a steady state model of the motor, a transformation of the motor states is
used. This transformation takes the states described in the stator fixed reference frame
xdq and transform these to an arbitrary frame. In this case the arbitrary frame rotates
with the frequency of the supply voltage e . The new states are denoted xedq . The
transformation is given by xdq = T(e )xedq , where e is the angle between the stator
fixed frame and the arbitrary rotating frame. T is given by
cos(e ) sin(e )
T(e ) = , (7.2)
sin(e ) cos(e )
where e is a function of time. Using this transformation to transform vsdq , isdq , and
imdq , a new model is obtained with the property that all states are constant during steady
state operation. Setting the derivatives of the states equal to zero, the following steady
state model of the motor is obtained,
e
e L0s Isq = (Rs + Rr0 )Isd
e e
+ (Rr0 Imd e
zp r L0m Imq e
) + Vsd
0 e 0 e 0 e 0 e e
e Ls Isd = (Rs + Rr )Isq + (Rr Imq + zp r Lm Imd ) + Vsq
e L0m Imq
e e
= (Rr0 Imd e
zp r L0m Imq ) + Rr0 Isd
e
0 e 0 e 0 e 0 e
e Lm Imd = (Rr Imq + zp r Lm Imd ) + Rr Isq .
e
Defining Isd = 0 in the above model, meaning that the rotating reference frame is
e
aligned with the current Isq , the final steady state model of the motor is obtained,
e L0s Isq
e
= (Rr0 Imd
e
zp r L0m Imq
e e
) + Vsd
0 = (Rs + Rr )Isq + (Rr Imq + zp r L0m Imd
0 e 0 e e e
) + Vsq
0 e 0 e 0 e (7.3)
e Lm Imq = (Rr Imd zp r Lm Imq )
e L0m Imd
e
= (Rr0 Imqe
+ zp r L0m Imde
) + Rr0 Isq
e
.
The torque expression of the motor is given in (3.11) and repeated here for the sake of
convenience,
3
Te = zp L0m (imd isq imq isd ) .
2
Using the transformation T, given in (7.2), a torque expression of the new variables is
obtained. The expression becomes
3 e e
Te = zp L0m Imd Isq . (7.4)
2
Equations (7.3) and (7.4) form the model of the induction motor during steady state op-
e e e
erations. However, the electrical quantities Vsd , Vsq , and Isq are not directly measurable.
Therefore, a connection between these quantities and the measurable quantities must be
established. The measurable quantities are the RMS values of the supply voltage Vrms
and current Irms , the supply frequency e and the electrical angle between the voltage
155
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
and current . The angle is in this case defined as the angle between the voltage vec-
T
e T
e
tor Vtdq = Vtde Vtqe , and the current vector Ietdq = 0 Itq , where subscript t
denotes quantities available at the terminals of the motor. The Euclidean length of the
e
vectors Vtdq and Ietdq is connected to the RMS values of the measurable electrical quan-
e
tities by |Vtdq | = 2Vrms and |Ietdq | = 2Irms . Using the RMS values Vrms and Irms
and the angle the voltages Vtde and Vtqe , and the current Itq
e
can be calculated using
Vtde = 2Vrms sin() Vtqe = 2Vrms cos() e
Itq = 2Irms . (7.5)
In the case of a Y-connected motor the electrical quantities at the motor terminals Vtde ,
Vtqe , and Itq
e e
and the quantities Vsd e
, Vsq e
, and Isq in (7.3) are equivalent. However, in
the case of a -connected motor the transformation matrices Bv and Ci have to be
considered. These matrices are defined in Section 3.2.3. It can be shown that
1
Bv = 3T C1i = T ,
3
where T is a rotation matrix. Using these expressions in (7.1) to obtain the steady
state model, it is seen that (7.3) can be used to model the steady state operation of the
motor when it is connected in a -connection if the following scalings of the electrical
quantities are used,
1 e
Vtde = 3Vtde Vtqe = 3Vtqe e
Isq = Itq . (7.6)
3
0 = Te Br fT (Qp , r ) . (7.7)
The hydraulic parts of the pump is considered as the parts involved with the energy
transformation from mechanical to hydraulic energy. In the model derived in Section
3.3 this energy transformation is described by two maps given by
156
Section 7.1: Steady State Model of the System
where Tp and Hp are the load torque and pressure produced by the pump respectively.
The arguments of the maps fT and fH in (7.8) are the volume flow Qp and the speed of
the pump r . The maps fT and fH are given by,
fH (Qp , r ) = g ah2 Q2p + ah1 Qp r + ah0 r2
fT (Qp , r ) = at2 Q2p + at1 Qp r + at0 r2 .
A general model of a hydraulic application of a centrifugal pump is given in Sec-
tion 6.3.1. Here it is argued that such a model, in most cases, can be described by the
following state space model,
xQ = fA (xQ , Hp , d)
Qp = hA (xQ ) ,
where xQ is the state vector of the application model, Hp and Qp are the pressure and
volume flow of the centrifugal pump, and d is a vector containing some unknown input
signals to the application. If it is assumed that the derivative of the states xQ equals zero
during steady state operation, the steady state model becomes,
0= fA (xQ , Hp , d)
(7.9)
Qp = hA (xQ ) .
157
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
c1 : e L0s Isq
e
= A1 + Vsd e
c5 : 0 = Te Br Tp
c6 : 0 = fA (xQ , Hp , uQ )
c7 : A1 = Rr0 Imde
zp r L0m Imqe
0 e 0 e
c8 : A2 = Rr Imq + zep er Lm Imd (7.10)
c9 : Te = 23 zp Lm Isq Imd
c10 : Hp = ah2 Q2p + ah1 Qp r + ah0 r2
c11 : Tp = at2 Q2p + at1 Qp r + at0 r2
c12 : Qp = hA (xQ )
e
c13 : Y1 = Isq
c14 : Y2 = e
c15 : Y3 = Hp
c16 : Y4 = Qp
tions in (7.3) are split into six equations to avoid using redundant structural information
in the SA.
The graph representation of the constraints in (7.10) is shown in Table 7.1. From
this table it is seen that the set of constaints Ce = {c1 , c2 , c3 , c4 , c7 , c9 , c13 , c14 } forms a
match on the unknown variables contained in these constraints. Therefore, this part can
be considered as a subsystem, with two outputs r and Te . These outputs are denoted
the connecting variables between this subsystem and the remaining parts of the system.
The set of constraints Ce is in the following denoted the first subsystem, whereas the
remaining set of constraints Cm = {c5 , c6 , c10 , c11 , c12 , c15 , c16 } is denoted the second
subsystem. This division of the system is equivalent to the division described in Chapter
6.
In Section 7.1.3 it is argued that the faults considered in this section only affect the
hydraulic part of the pump. The hydraulic part of the pump is described by fH , fT and
the expression of the flow measurement. These expressions are given by the constraints
c10 , c11 , and c16 , which are included in Cm only. Therefore, only the second subsystem
Cm is affected by the faults considered in this chapter, meaning that only this part must
158
Section 7.2: Structural Analysis
Table 7.1: The structure table of the centrifugal pump system under steady state condi-
tions. The symbols show a matching for the first part of the system.
Known Unknown
e Ve Y
Y1 Y2 Vsd e Ie A A Ie
sq 3 Y4 xQ uQ Tp Hp Qp Te r Imd mq 1 2 sq e
c6 1 1 1
c12 1 1
c5 1 1 1
c11 1 1 1
c10 1 1 1
c15 1 1
c16 1 1
c8 1 1 1 1
c9 1 1
c7 1 1 1
c4 1 1 1
c3 1 1
c1 1 1 1
c2 1 1
c13 1
c14 1
159
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
Beside these constraints two extra constraints are added to model the estimation of the
connecting variables r and Te . These constraints are given by
c17 : Te = Te
(7.13)
c18 : r = r .
Table 7.2 is a representation of the graph describing the structure of the second subsys-
tem. In this table the column at the right hand side is added to show the connection
between the faults in the system and the constraints. Using SA three minimal over-
Table 7.2: The structure table of the second system. The first two columns describe
the structural connection between the constraints and the known and unknown variables
respectively. Whereas the last column describes the connection between the constraints
and the faults in the system.
Known Unknown Faults
Y3 Y4 Te r xQ uQ Tp Hp Qp Te r Kf Kl B fc fd
c6 1 1 1
c12 1 1
c5 1 1 1
c11 1 1 1 1 1 1
c10 1 1 1 1 1 1
c15 1 1
c16 1 1
c17 1 1
c18 1 1 1
constraint subsystems are identified. These are given by the following three sets,
The connections between these sets and the faults in the system are given by
Cm1 : {Kl , B, fc , fd }
Cm2 : {Kf , B, fc , fd } (7.15)
Cm3 : {Kf , Kl , fc , fd } .
160
Section 7.2: Structural Analysis
where pi is the parameter vector and ypi is the measurement vector. In this expression
both api () and fpi (y) contain polynomial functions. In the following the parameter
and measurement vectors pi and ypi , and the polynomial vector functions fpi and api
are given for each set of constraints Cm1 , Cm2 , and Cm3 .
The set of constraints Cm1 contains the following parameters and measurements de-
scribed on vector form,
T T
p1 = at2 at1 at0 yp1 = Te r Y4 .
The set of constraints Cm2 contains the following parameters and measurements de-
scribed on vector form,
T T
p2 = ah2 ah1 ah0 at2 at1 at0 yp2 = Te r Y3 .
161
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
ap2 ( p2 ) =
a2h2
a2t2
ah2 at1 ah1 at2 a2 + 2a2 at0 2ah2 ah0 at2
h1 h2 ,
a1 () + a2 ()
2ah2 at2 at0 + ah2 a2t1 2ah0 a2t2 ah1 at2 at1
2ah2 at2
where
and
2 2
B r 2BTe r + Te2
Y32
B 3
T 2
fp1 (yp1 ) = r e r .
4
r
Y3 r2
Y3 Te + BY3 r
162
Section 7.3: The Robust FDI Algorithm
Moreover, the density of the liquid is not constant due to temperature changes and im-
purities in the liquid. This density affects the value of the parameters in the hydraulic
model of the pump given by (3.29). Therefore, parameter variations must be taken into
account, for the developed algorithms to be robust and thereby usable in real applica-
tions.
Definition 7.3.1 The set of possible residual values R has the following properties,
If the system is not affected by faults i.e. e = 0 then 0 R.
If 0
/ R then there are faults in the system i.e. e 6= 0
where e is a given fault vector and R R is a connected set.
Now let r = g(y, ) be an ARR describing the behaviour of a given system, where y
contains the measurements and contains the parameters in the ARR. For this ARR r
only equals 0 in the no fault case, if the structure and parameters of the system are known
exactly. When, for such a system, parameter uncertainties are taken into account the set
of possible residuals R defined in Definition 7.3.1 is given by the following lemma,
163
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
This is a well known fact from mathematical analysis and the proof can be found in
(Apostol, 1974). Using this lemma the boundaries of R can be calculated. Unfortunately
it is not straightforward to calculate these maximum and minimum values. However,
in the case treated in this work the function g can be separated into a function of the
parameters and a function of the measurements i.e. g(y, ) = f (y)T a(). When this is
the case it is easy to see that one set of boundaries on the residual set R is given by
k
X
r rmax = max (ai ()fi (y))
i=1
k
X
r rmin = min (ai ()fi (y)) .
i=1
Lemma 7.3.2 If r = f (y)T a() where a and f are two vector functions, which are
continuous on respectively and Y, then an upper (lower) boundary of r (r) is given
by
X X
r rmax = ri,max r rmin = ri,min ,
i i
where
fi (y) max (ai ()) if fi (y) > 0
ri,max =
fi (y) min (ai ()) otherwise
fi (y) min (ai ()) if fi (y) > 0
ri,min =
fi (y) max (ai ()) otherwise .
The upper and lower boundaries, found using this lemma, are fast and easy to calculate
online, as the hard part of the calculations can be done offline. Unfortunately, the found
upper and lower boundaries are in general not very tight. However, if the residual ex-
pression r = g(y, ) is affine with respect to the parameters an exact solution exists.
If g is affine with respect to the parameters it can be rewritten to be come,
For an expression of this form the upper and lower boundaries can be found on an
interval set using interval algebra (Boukhris et al., 1998). Using this approach the
exact maximum and minimum values of (7.17) can be calculated, in the no fault case,
using the following lemma.
164
Section 7.3: The Robust FDI Algorithm
Lemma 7.3.3 For the parameter affine system (7.17) the maximum and minimum values
of the residual set R can be calculated exactly, and are given by
where 0 = 12 ( + ).
This is a well-known fact from interval analysis (Boukhris et al., 1998; Moore, 1979).
In the following, Lemma 7.3.2 and 7.3.3 are used to calculate an upper and a lower
boundary of the residual expressions found in the previous section.
All the residual expressions derived in Section 7.2.3 are on the form
r = f (y)T a() ,
where both a and f are vector functions with only polynomial nonlinearities in and
y. As this is a nonlinear expression, Lemma 7.3.2 could be used to find upper and
lower boundaries of the residual set. However, these upper and lower boundaries are in
general far from the real maximum and minimum values of R, which is also the case in
centrifugal pumps. Therefore, the linear dependency between the parameters is extracted
using a first order Taylor Series expansion of a(). This Taylor Series expansion is given
by
a
a() = a( 0 ) + i ( 0 ) ( 0 ) + O( 0 ) .
When this expression is used in the residual equation the residual is calculated by the
sum of two terms r = rl + rO . These terms are given by
a
rl (y) = f (y)T a( 0 ) + f (y)T i
( 0 ) (7.18a)
rO (, y) = f (y)T O() , (7.18b)
165
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
y compute
1. At each new sample f (y), G0 (y) = f (y)T a( 0 ) and
T a ( )
Gl (y) = f (y) i 0 .
2. Compute the maximum and minimum values rl and rl using
Lemma (7.3.3).
3. Compute the boundaries rO,min and rO,max using Lemma 7.3.2.
4. Compute the boundaries of the residual set using,
r rmin = rl + rO,min
r rmax = rl + rO,max
Remark 7.3.1 This algorithm can calculate the boundaries of the residual set of one
single residual expression. However, parameter dependencies between residuals, when
more than one residual is considered, are not taken into account. Therefore, the identi-
fication of incipient faults can create problems in some cases.
2 rmax
+ yi r rO,max
rl
rl
r rO,min
rmin
1 yi
Figure 7.1: Illustration of the different boundary values in the proposed algorithm.
To the left of Fig. 7.1 the set of possible parameter values is shown. At each
measurement vector yi this set of parameter values is mapped into a set of possible
166
Section 7.4: Test Results
residual values. In the right-hand side figure of Fig. 7.1 this set is defined by r, r. Now
considering the developed algorithm. The residual set defined by rl , rl in Fig. 7.1 is
calculated in step 2, and rO,min , rO,max are calculated in step 3. These values sum up
to a set defined by rmin , rmax , which is in general conservative compared to the real
residual set defined by r, r. If r = 0 is not in between r, r or alternative not in between
rmin , rmax it is guaranteed that a fault has happened in the system.
vrms Te
irms Motor Interval rp1
e Equations Algorithm
cos() r based on
rp2
Hp ARR's rp3
Qp
Figure 7.2: The final detection algorithm. The first part consists of (7.11) and (7.12)
used for calculating the connecting variables. The second part consists of the interval
algorithm described in Section 7.3.
In this algorithm the first part consists of (7.11) and (7.12), which are used for cal-
culating the connecting variables r and Te . These calculations are based on the RMS
values of the supply voltage Vrms and current Irms , cos() where is the angle between
the supply voltage and current, and the frequency of the supply voltage e , see Section
7.1. The connection variables r and Te are then, together with the flow and pressure
measurement Qp and Hp , used as input to the second subsystem, which consists of the
residual generators.
This detection algorithm is tested on the same test setup as the algorithm described
in Chapter 6. Meaning that it is tested on a Grundfos 1.5 (KW ) CR5-10 pump placed
in a tank system as depicted in Fig. 7.3. In this tank system the valve V1 is used to
model disturbances in the system. Clogging inside the pump is modelled by the valve
Vc and dry running is modelled by closing V2 and opening V3 . Rub impact is modelled
by adding an extra force to the shaft and cavitation is modelled by closing valve V2
gradually. Leakage flow is modelled by opening Vl . In the tests shown here a low
167
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
tank
Hp
Vl
V1 V2
pump V3
Vc
pipe Qp
Te
shaft
M
motor
Figure 7.3: Sketch of the test setup. The measurements are the electrical quantities, the
differential pressure Hp delivered by the pump and the volume flow through the pump
Qp .
bandwidth but precise flow sensor is used. This flow sensor is chosen, as this type of
sensor is often used in real industrial applications.
First robustness with respect to the operating point is tested. In this test both the
position of the valve V1 and the speed of the pump are changed during operation. During
the test the valve is changed in three steps from maximum opened to almost closed. The
first step change is done at time 26.3 [sec] and the second step change is done at time
52.3 [sec]. The speed of the pump is changed between approximately 2030 and 2580
(rpm) each 13 [sec] during the test. The result of this test is shown in Fig. 7.4. In the
three lower figures the residual sets Rp1 to Rp3 are shown by their boundaries at each
time instant, and in the top figure the speed changes are shown. The test shows that
after each transient phase, zero is included in all the three residual sets, i.e. zero is in
between the residual boundaries for each residual set. This shows that the algorithm is
not capable of handling transient behaviour, which was expected. More important, it
shows that the algorithm is robust with respect to the operating point of the pump.
Figures 7.5 to 7.7 show test results concerning isolability of the five faults considered
in this chapter. All these tests are performed with V1 half opened and an angular speed
of approximately 2350 (rpm). In each of the figures the residual sets Rp1 to Rp3 are
shown by their boundaries in the top figure, and the decision Dp1 to Dp3 are shown
in the bottom figure. The tests show that all the five faults considered in this chapter
are detectable using the three residual sets. Moreover, comparing the decision signals
it is seen that all the fault signatures given by Dp1 , Dp2 , and Dp3 are distinguishable
except from the cavitation and dry running fault, i.e. fc and fd . This was expected as
168
Section 7.5: Discussion
Speed [rad/sec]
300
250
200
0 10 20 30 40 50 60 70
0.5
Rp1
0.5
0 10 20 30 40 50 60 70
0.02
Rp2
0.02
0 10 20 30 40 50 60 70
0.1
Rp3
0.1
0 10 20 30 40 50 60 70
time [sec]
Figure 7.4: Robustness test. The top figure shows the speed steps, and the three lower
figures show the residual boundaries for each of the three residual sets Rp1 to Rp3 .
7.5 Discussion
The topic of this chapter is fault detection in a centrifugal pump based on steady state
measurements only. Here the steady state measurements are Root Mean Square (RMS)
measurement of the voltage and current, the voltage frequency, the electrical angle be-
tween the voltage and current, and the pressure and volume flow of the pump. Three
residuals are derived form the steady state model of the pump, using Structural Analy-
sis and Analytical Redundant Relations (ARRs). The faults under consideration in this
chapter are only affecting the second subsystem. Hence the ARRs are only developed
for this part. The connecting variables between the two subsystems are calculated using
a steady state model of the motor.
Structural analysis has been used to analyse the system. As a result of this analysis
the system is divided into two cascade-connected subsystems simplifying the derivation
of the Analytical Redundancy Relation (ARR) considerably. The first subsystem con-
sists of the induction motor model, and the mechanical and hydraulic parts of the pump
169
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
0.4 Rp1
Rp2
0.2 Rp3
Residual sets
0
0.2
0.4
0 2 4 6 8 10 12 14 16 18 20
3
Dp1+2
2.5 Dp2+1
Dp3
Decision signals
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18 20
time [sec]
0.4 Rp1
Rp2
0.2 Rp3
Residual sets
0.2
0.4
0 2 4 6 8 10 12 14 16 18 20
3
Dp1+2
2.5 Dp2+1
Dp3
Decision signals
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18 20
time [sec]
Figure 7.5: Test results. The top figures show the residual boundaries for each of the
three residual sets R1 , R2 and R3 . The bottom figures show the decision signals D1 ,
D2 , and D3 .
170
Section 7.5: Discussion
1.5 Rp1
Rp2
1
Rp3
Residual sets
0.5
0.5
1.5
0 2 4 6 8 10 12 14 16 18
3
Dp1+2
2.5 Dp2+1
Dp3
Decision signals
1.5
0.5
0
0 2 4 6 8 10 12 14 16 18
time [sec]
1.5 Rp1
Rp2
1
Rp3
Residual sets
0.5
0.5
1.5
0 2 4 6 8 10 12 14 16
3
Dp1+2
2.5 Dp2+1
Dp3
Decision signals
1.5
0.5
0
0 2 4 6 8 10 12 14 16
time [sec]
Figure 7.6: Test results. The top figures show the residual boundaries for each of the
three residual sets R1 , R2 and R3 . The bottom figures show the decision signals D1 ,
D2 , and D3 .
171
Chapter 7: FDI on the Centrifugal Pump: A Steady State Solution
2 Rp1
R
p2
Rp3
1
Residual sets
0
0 5 10 15 20
3
Dp1+2
2.5 Dp2+1
Dp3
Decision signals
1.5
0.5
0
0 5 10 15 20
time [sec]
Figure 7.7: Test results showing detection of the fault fd dry running. The top figures
show the residual boundaries for each of the three residual sets R1 , R2 and R3 . The
bottom figures show the decision signals D1 , D2 , and D3 .
form the second subsystem. The approach of dividing the system into two cascade-
connected subsystems was also utilized in Chapter 6. Parameter variations are only
considered for the second subsystem, with the drawback that the algorithm can handle
only parameter variations in the hydraulic part of the pump. However, the possibility
of handling parameter variations in the hydraulic part can be used to obtain a simple
and logical way of setting alarm levels for a user of the system, as these can be defined
directly on pump curves, as those shown in Section 3.3.5.
The residuals obtained from the ARRs are made robust with respect to parame-
ter variations in the centrifugal pump model by using the set-valued approach. It is
shown that linearization of the parameter function can be used to calculate relatively
tight boundaries of the residual in the centrifugal pump case. Using the set-valued ap-
proach the set of possible residual values are changed as a function of the operating
point of the pump. This could be compared to a residual with an adaptive threshold. The
presented method has the advantage of connecting the physics of the pump and the set
of residuals in a straightforward manner.
172
Chapter 8
Conclusion and
Recommendations
In this thesis different aspects of Fault Detection and Identification (FDI) in centrifugal
pumps were considered. Special focus was put on the robustness of the FDI algorithms.
In this connection an analysis method for analysing robustness in signal-based fault
detection schemes was proposed. In most model-based fault detection schemes, robust-
ness considerations are a part of the design. Therefore, the possibilities of using these
approaches on the centrifugal pump were investigated too, ending up with three new FDI
algorithms. Here, a small example of connecting these algorithms into one FDI scheme
is given. This example is followed by a conclusion and a number of recommendations
for further research.
173
Chapter 8: Conclusion and Recommendations
6. Dry running.
In chapter 5 the detection and identification of the inter-turn short circuit were consid-
ered, and in Chapters 6 and 7 the remaining five faults were handled. The difference
between the two algorithms in Chapters 6 and 7 lies in their ability to handle transient
behaviour in the system. The algorithm derived in Chapter 6 was based on a dynamic
description of the pump and was therefore able to handle transient behaviour. The al-
gorithm derived in Chapter 7 was on the other hand based on a steady state model and
therefore had inherent problems during transient phases.
Composing the algorithm derived in Chapter 5 with the FDI part of the algorithm
derived in Chapter 6, the final FDI scheme is obtained. This FDI scheme is capable of
detecting and identifying the 6 different faults in the centrifugal pump. The composition
of the two algorithms is shown in Fig. 8.1.
Real
world
Tl
Pump Q,H
Vs Motor Is
Software
algorithm
D1
Te Residual Mech. and
D2
observers hyd. faults.
D3
Adaptive
fp1
observers
fp2 Stator faults.
fp3
Figure 8.1: The composition of the adaptive observer, designed for inter-turn short cir-
cuit detection, and residual observers, designed for fault detection and identification in
the mechanical and hydraulic part of the centrifugal pump.
In Fig. 8.1 the block denoted adaptive observer, contains the observer used for de-
tecting stator faults. The output from this observer is the estimates of the electrical states,
the speed, and the inter-turn short circuit. The estimates of the electrical states are used
for calculating the torque, meaning that the adaptive observer is capable of estimating
both the speed, the torque, and the inter-turn short circuit fault.
The inputs to the block, denoted residual observers in Fig. 8.1, are the measured
pressure, the measured volume flow, and the estimates of the speed and torque. The
output of the residual observer block is the three decision signals described in Chapter
6, meaning that the information of the mechanical and hydraulic faults are available.
In Chapter 7 residual generators, designed using the steady state model of the me-
chanical and hydraulic parts of the centrifugal pump, are described. These residual gen-
erators could also be used for residual generation in the block, denoted residual observer
174
Section 8.2: Conclusion
in Fig. 8.1. However, by using these residual generators, problems will arise during
transient phases.
8.2 Conclusion
The following conclusions are drawn on the accomplishments and contributions of the
thesis:
The dynamics of the centrifugal pump were considered in the development of the
centrifugal pump model in Chapter 3. Two assumptions were given, under which
algebraic maps can be used for describing the energy conversion from mechani-
cal to hydraulic energy, and the pressure of the centrifugal pump. Moreover, the
dynamics of the hydraulic part can be described by an added mass to the impeller,
which is corresponding to the mass of the liquid inside the impeller. This means
that, when the mechanical and hydraulic part of the centrifugal pump is consid-
ered, the dynamics of the pump is described by a first order system. The algebraic
maps describing the energy conversion and the pressure generation are polyno-
mial of second order. This makes the model of the system a nonlinear lumped
parameter model. As the model is a lumped parameter model, it is well suited for
use in model-based designs, both when it comes to control and to FDI.
The most used methods for FDI in centrifugal pumps are, according to the state
of the art analysis, presented in Chapter 2, based on the signal-based approaches.
Normally robustness is not considered when FDI algorithms are derived using
signal-based approaches. Therefore, it was found that a method, which can be
used for robustness analysis, was needed. In Chapter 4 a combination of the Fail-
ure Mode and Effect Analysis (FMEA) and Fault Propagation Analysis (FPA) is
proposed for the robustness analysis of signal-based FDI algorithms. The pro-
posed method can be used in the design, as well as in the analysis. In addition
to the robustness analysis method, an algorithm for automating one of the manual
steps in the FPA was proposed. Using the improvements of the FMEA and FPA,
the proposed analysis method was used to analyse different sensor configurations
on the centrifugal pump. Also it was used in the design of a detection scheme,
based on one particular sensor configuration. Test results on an industrial test-
bench showed that 3 out of 7 chosen faults could be detected in a robust manner
with this sensor configuration. This showed the usability of the proposed analysis
and design method. However, the result obtained in Chapters 6 and 7 showed that,
using the model-based approach, 5 different faults were detected using the same
sensor configuration. This shows the drawback of the signal-based approaches
when used on centrifugal pumps.
One of the faults considered in this work was the inter-turn short circuits in the
windings of the induction motor driving the pump. In the state of the art analysis
175
Chapter 8: Conclusion and Recommendations
in Chapter 2, it was argued, that this fault is often the initiator of a stator burnout.
To understand the nature of the inter-turn short circuit fault, a dynamic model
of the fault was derived in Chapter 5. A model was established for both Y- and
-connected induction motors. One remarkable result was shown by this model,
namely that torque ripple should not be expected when an inter-turn short circuit
has occurred. That is, if the motor is supplied with a balanced sinusoidal supply
voltage, and a constant torque load. This fact was indirectly shown in the test,
presented in the end of Chapter 5, as the level of oscillations in the real speed did
not change when introducing the inter-turn short circuit fault.
In Chapter 5, the model of the induction motor, including an inter-turn short cir-
cuit, was used in the derivation of an adaptive observer. In real applications, only
the electrical quantities are normally available. Therefore, the derived adaptive
observer was only using these quantities. The obtained observer was capable of
estimating an inter-turn short circuit fault in one of the phases, the electrical states,
and the speed of the induction motor simultaneously. The design of the observer
was based on a dynamic model of the induction motor, with the speed assumed
constant. Therefore, the observer was capable of detecting the faults during tran-
sient behaviour, when the constant speed assumption was not violated too much.
This makes the algorithm very useful in real applications. Especially for induction
motors supplied with a frequency converter, as transients are expected to occur
frequently in these applications. The drawback of the algorithm is the need for
the motor parameters, which are not always known in real life applications.
In Chapter 3 it was shown that the model of the centrifugal pump is highly non-
linear, meaning that model-based approaches, based on nonlinear models, should
be considered when the operating point of the pump is changed frequently or is
unknown. This is actually the case in many centrifugal pump applications. Here,
Structural Analysis (SA) was chosen as the first step in the derivation of residual
generators. The only straightforward way to use the results of the SA, in order
to obtain residual generators, is the derivations of Analytical Redundant Relations
(ARRs). These ARRs can, and in general do, contain derivatives of the measured
signals. To overcome this problem, a novel realization approach was proposed in
Chapter 6. With this approach, it is possible to obtain a nonlinear state space de-
scription from the results of the SA. The obtained state space description does not
include unknown variables, except for the states of the system. The approach was
tested on three applications; a satellite case borrowed from (De Persis and Isidori,
2001), the induction motor, and the hydraulic part of the centrifugal pump. The
realization method is based on a solution to a partial differential equation, which
is the main problem with this approach. This is, however, a known problem con-
cerning nonlinear state transformations on general nonlinear systems, as it is the
case in the proposed realization approach.
Based on the results obtained using SA and the realization approach developed in
176
Section 8.3: Recommendations
Chapter 6, three residual observers and an ARR were derived for the centrifugal
pump. Tests, on a test-bench developed particularly for this purpose, were per-
formed, showing the capabilities of the observers and the ARR. The tests showed
that the observers and the ARR were robust, with respect to changes in the operat-
ing point of the pump. Moreover, the tests showed that 5 different faults, affecting
the hydraulic and mechanical parts of the pump, were detectable with this ap-
proach. This shows the superiority of the model-based approach compared to the
signal-based approach, when it comes to robustness.
In Chapter 6, the algorithm, developed for detecting faults in the mechanical and
hydraulic part of the centrifugal pump, was based on a dynamic model of the sys-
tem. This means that the involved signals were expected to have a bandwidth
covering the dynamics of the system. These high bandwidth signals are often not
available in real life centrifugal pump applications. However, low bandwidth sig-
nals are often available. Therefore, an algorithm based on a steady state model
of the centrifugal pump was proposed in Chapter 7. This algorithm was shown
to be able to detect 5 different faults in the mechanical and hydraulic parts of
the centrifugal pump. The algorithm was made robust with respect to parameter
uncertainties, using a set-valued approach. The need for only low bandwidth sen-
sors and the robustness considerations, makes the algorithm appropriate for im-
plementation in cost sensitive applications, where the necessary sensors already
are available.
8.3 Recommendations
The following topics are not covered in this thesis, but it is believed that future investi-
gations could be beneficial.
In Chapter 5 an adaptive observer is designed for simultaneous estimation of the
electrical states, the speed, and the inter-turn short circuit of the induction mo-
tor. In the design proposed here, the parameters of the motor are assumed known.
However, an adaptive observer is used. Therefore it should be possible to find
a transformation of the induction motor model, making it possible for the adap-
tive observer to adapt to the motor parameters. If this is possible, the proposed
algorithm becomes parameter independent.
In Chapter 5 also a Linear Matrix Inequality (LMI) approach was used for both
analysis and synthesis of the observer gain in the proposed adaptive observer. It
could be interesting to include the concept of LMI regions in the proposed analysis
and synthesis. Hereby, it would be possible to state demands on the damping ratio
in the system, as well as the convergence rate.
In Chapter 6 a new concept, for realization of subsystems identified using Struc-
tural Analysis, was presented. This concept should be further investigated. Es-
177
Chapter 8: Conclusion and Recommendations
pecially Assumption 6.2.3, stating the condition for the existence of a state space
transformation, should be considered.
In Chapter 7 a detection scheme, based on the steady state model of the pump,
was proposed. Because of the steady state model, transient phases can force the
residual to be different from zero, even in the no fault cases. Therefore, means for
detecting these transient conditions, must be established before implementing the
algorithm in real life applications.
A set-valued approach was proposed for obtaining robustness in the residual gen-
eration in Chapter 7. This approach utilizes interval models. Extending the set-
valued approach, to include interval models of each fault case, it might be possible
to state which fault has happened in a robust manner. Hereby robustness is intro-
duced in the identification of incipient faults.
178
Bibliography
Neil Arthur and Jim Penman. Induction machine condition monitoring with higher
order spectra. IEEE Transaction on Industrial Electronics, Vol. 47, No. 5, 2000.
Antonio Baldassarre, Maurizio De Lucia, and Paolo Nesi. Real-time detection of cav-
itation for hydraulic turbomachines. Real-Time Imaging 4, Page 403 - 416, 1998.
Michele Basseville and Igor V. Nikiforov. Detection of Abrupt Changes - Theory and
Application. www.irisa.fr/sigma2/kniga/, 1998.
Alberto Bellini, Fiorenzo Filippetti, Giovanni Franceschini, Carla Tassoni, and Ger-
ald B. Kliman. Quantitative evaluation of induction motor broken bars by means of
electrical signature analysis. IEEE Transaction on Industry Applications, Vol. 37, No.
5, 2001.
Sren Abildsten Bgh. Fault Tolerant Control Systems - a Development Method and
Real-Life Case Study. PhD thesis, Aalborg University, Department of Control Engi-
neering, 1997.
179
Bibliography
Pter Bka and Gbor Halsz. Dynamic behaviour of centrifugal pumps. Proceedings
of the XXIst IAHR Symposium on Hydraulic Machinery and Systems, Lausanne, 2002.
Mogens Blanke. Consistent design of dependable control systems. Control Eng. Prac-
tice, Vol. 4, No. 9, 1996.
Mogens Blanke, Michel Kinnaert, Jan Lunze, and Marcel Starowiecki. Diagnosis and
Fault-Tolerant Control. Springer Verlag, 2003. ISBN 3-540-01056-4.
Julien Le Bleu Jr. and Ming Xu. Vibration monitoring of sealess pumps using spike
energy. Sound and Vibration, Vol. 29, No. 12, 1995.
Austin H. Bonnett and George C. Soukup. Cause and analysis of stator and rotor
failures in three-phase squirrel-cage induction motors. IEEE Transaction on Industry
Applications, Vol. 28, No. 4, 1992.
A. Boukhris, D. Mandel, and J. Ragot. Data validation using interval algebra. UKACC
International Conference on Control, CONTROL98, 1998.
Fernando Briz, Michael W. Degner, Antonio Zamarron, and Juan M. Guerrero. Online
stator winding fault diagnosis in inverter-fed ac machines using high-frequency signal
injection. IEEE Transactions on Industry Applications, Vol. 39, No. 4., 2003.
Cash, M. A., Habetler, T. G., and Kliman, G. B. Insulation failure prediction in in-
duction machines using line-neutral voltages. IEEE Industry Applications Conference,
Thirty-Second IAS Annual Meeting, IAS 97, 1997.
Jie Chen and R. J. Patton. Robust Model-Based Fault Diagnosis for Dynamic Systems.
Kluwer Academic Publishers, 1999. ISBN 0-7923-8411-3.
Young Man Cho and Rajesh Rajamani. A systematic approach to adaptive observer
synthesis for nonlinear systems. IEEE Transactions on Automatic Control, Vol. 42, No.
4, 1997.
Tommy W. S. Chow and Hong-Zhou Tan. Hos-based nonparametric and parametric
methodologies for machine fault detection. IEEE Transaction on Industrial Electron-
ics, Vol. 47, No. 5, 2000.
David Cox, John Little, and Donal OShea. Ideals, Varieties, and Algorithms, An
introduction to computational algebraic geometry and comutative algebra. Springer-
Verlag Inc., 1997. ISBN 0-387-94680-2.
Sergio M. A. Cruz and A. J. Marques Cardoso. Stator winding fault diagnosis in three-
phase synchronous and asynchronous motors, by the extended parks vector approach.
IEEE Transaction on Industrial Applications, Vol. 37, No. 5, 2001.
M. Cudina. Detection of cavitation phenomenon in centrifugal pump using audible
sound. Mechanical Systems and Signal Processing, Vol. 17, No. 6, 2003.
180
Bibliography
Francesco Cupertino, Elisabetta de Vanna, Luigi Salvatore, and Silvio Stasi. Analysis
techniques for detection of im broken rotor bars after supply disconnection. IEEE
Transactions on Industry Applications, Vol. 40, No. 2., 2004.
Tracy Dalton, Ron J. Patton, and J. Chen. Application of eigenstructure assignment to
robust residual design for FDI. UKACC International Conference on Control, CON-
TROL 96, 1996.
Claudio De Persis and Alberto Isidori. A geometric approach to nonlinear fault detec-
tion and isolation. IEEE Transaction on Automatic Control, Vol. 46 No. 6, 2001.
Diego Del Gobbo, Marcello Napolitano, Parviz Famouri, and Mario Innocenti. Expe-
rimantal application of extended kalman filtering for sensor valitation. IEEE Transac-
tions on Control Systems Technology, Vol 9. No 2., 2001.
Levent Eren and Michael J. Devaney. Motor bearing damage detection via wavelet
analysis of the starting current transient. IEEE Instrumentation and Measurement Tech-
nology Conference, 2001.
P. M. Frank and X. Ding. Survey of robust residual generation and evaluation methods
in observer-based fault detection systems. Journal of Proces Control, Vol. 7, No. 6,
1997.
Jens Friedrichs and Gnter Kosyna. Rotating cavitation in a centrifugal pump impeller
of low specific speed. Transactions of the ASME, Journal of Fluid Engineering, Vol.
124, No. 2, 2002.
Erik Frisk and Jan slund. Lowering orders of derivatives in non-linear consistency
relations - theory and simulation examples. Technical Report LiTH-R-2547, Dept. of
Electrical Engineering, Linkping University, Sweden, 2003.
Pablo Garca, Fernando Briz, Michael W. Degner, and Alberto B. Diez. Diagnostics
of induction machines using the zero sequence voltage. Proceedings of the IEEE 39th
IAS Annual Meeting, Seattle, Washington, 2004.
E. Alcorta Garcia and P. M. Frank. Deterministic nonlinear observer-based approaches
to fault diagnosis: A survey. Control Eng. practice, Vol. 5 No. 5, 1997.
C. Gerada, K. J. Bradley, M. Sumner, P. Wheeler, S. Pickering, J. Clare, C. Whitley,
and G. Towers. The implications of winding faults in induction motor drives. Proceed-
ings of the IEEE 39th IAS Annual Meeting, Seattle, Washington, 2004.
Janos J. Gertler. Fault Detection and Diagnosis in Engineering Systems. Marcel
Dekker, Inc., 1998. ISBN 0-8247-9427-3.
Jan Tommy Gravdahl and Olav Egeland. Centrifugal compressor surge and speed con-
trol. IEEE Transactions on Control System Technology, vol. 7, No. 5, 1999.
181
Bibliography
Janati Idrissi, Olivier Adrot, and Jose Ragot. Multi-fault detection of systems with
bounded uncertainties. Proceedings of the 40th IEEE Conference on Decision and
Control, Orlando, Florida USA, 2001.
R. Isermann and P. Balle. Trends in the application of model-based fault detection and
diagnosis of technical processes. Control Eng. Practice, vol 5, No. 5, 1997.
Alberto Isidori. Nonlinear Control Systems, third edition. Springer-Verlag, 1995. ISBN
3-540-19916-0.
Roozbeh Izadi-Zamanabadi. Structural analysis approach for fault diagnosis and dis-
turbance decoupling. Technical report, Institute of Electronic Systems, Aalborg Uni-
versitet, Denmark, 2001.
Bin Jiang and Marcel Staroswiecki. Adaptive observer design for robust fault estima-
tion. International Journal of Systems Science, Vol. 33, No. 9, 2002.
Gojko M. Joksimovic and Jim Penman. The detection of inter-turn short circuits in the
stator windings of operating motors. IEEE Transaction on Industrial Electronics, Vol.
47, No. 5, 2000.
Rikke Bille Jrgensen. Development and Test of Methods for Fault Detection and
Isolation. PhD thesis, Aalborg University, Department of Control Engineering, 1995.
182
Bibliography
Carsten Skovmose Kallese, Henrik Rasmussen, Pierre Vadstrup, and Roozbeh Izadi-
Zamanabadi. Estimation of stator winding faults in induction motors using an adaptive
observer scheme. Proceedings of the IEEE 39th IAS Annual Meeting, Seattle, Wash-
ington, 2004c.
Dragica Kostic-Perovic, Muslum Arkan, and Peter Unsworth. Induction motor fault
detection by space vector angular fluctuation. Proceedings of the IEEE 35th IAS Annual
Meeting, 2000.
Sang-Bin Lee, Rangarajan M. Tallam, and Thomas G. Habetler. A robust, on-line turn-
fault detection technique for induction machines based on monitoring the sequence
component impedance matrix. IEEE Transactions on Power Electronics, Vol. 18, No.
3, 2003.
183
Bibliography
184
Bibliography
185
Bibliography
A. J. Stepanoff. Flow Pumps, Design and Application. John Wiley & sons, 1957. ISBN
0 471 82137 3.
Jakob Stoustrup and Henrik Niemann. Fault detection for nonlinear systems - a stan-
dard problem approach. Proceedings of the 37th IEEE Conference on Decision &
Control., 1998.
Jakob Stoustrup and Henrik Niemann. Fault estimation - a standard problem approach.
International Journal of Robust and Nonlinear Control, Vol. 12, No. 8, 2002.
Karl strm, Pedro Albertos, Mogens Blanke, Alberto Isidori, Walter Schaufelberger,
and Ricardo Sanz. Control of Complex Systems. Springer-Verlag, 2001. ISBN 1-
85233-324-3.
Charles R. Sullivan and Seth R. Sanders. Models for induction machines with magnetic
saturation of the main flux path. IEEE Transactions on Industrial Applications, Vol 31,
No. 4, 1995.
Jesper Sandberg Thomsen. A fault tolerant electronic steering system for a fork lift
truck. Technical report, Institute of Electronic Systems, Aalborg Universitet, Denmark,
2000.
Sebastian Tornill, Teresa Escobet, and Vicenc Puig. Fault detection using interval
models. SafeProcess 2000, IFAC, Budapest, 2000.
Frederik C. Trutt, Joseph Sottile, and Jeffery L. Kohler. Online condition monitoring
of induction motors. IEEE Transactions on Industry Applications, Vol. 38, No. 6, 2002.
186
Bibliography
187
Bibliography
188
Appendix A
This appendix contains tables describing the fault modes of the pump system. The pump
system is in Chapter 4 devided into 7 components. These are,
Shaft.
In each Section of this appendix one of the above components is analysed. This means
that a FMEA table is given and that the disturbing events affecting the component are
listed. The FMEA tables contain the following information; the name of the fault mode,
the causes of the fault, and the the effect of the fault. Here, the general components,
forming a centrifugal pump, are analysed. This means that no failure information of a
particular product is used. This information is necessary to include risk assessment and
frequency of the faults in the FMEA, and are therefore omitted here. The tables and
matrices in this appendix are the results of the FMEA described in Chapter 4.
189
Chapter A: FMEA Tables Describing Faults in the System
Figure A.1: The input/output structure of the electrical part of the induction motor.
fem3 Short circuit between Moisture or water in the eem,i3 Oscillations in the
windings in the motor. motor. length of the pack transform
current.
Mechanical stress.
High motor temperature. eem,T 1 Increased
temperature inside the motor.
wear-out failure on the
inter-turn insulation.
fem4 Short circuit to ground. Moisture or water in the eem,i3 Oscillations in the
motor. length of the pack transform
current.
Mechanical stress.
High motor temperature. eem,i4 Unbalanced stator
current.
wear-out failure on the turn
to ground insulation. eem,T 1 Increased
temperature inside the motor.
eem,t3 1st harmonic torque
oscillations.
190
Section A.1: Electrical part of the induction motor
Parts
fem6 Eccentric air gab due to Poor production. eem,i4 Unbalanced stator
bend or misaligned motor current.
Mechanical stress.
shaft. eem,t3 1st harmonic torque
oscillations.
The propagation matrix representation of the component is given by the following logi-
cal expression,
eem Aem em
fem fem + Ady edy (A.1)
where the fault vector fem and the effect vector eem equals,
0
T
fem = fem1 fem2 fem3 fem4 fem5 fem6
eem = eem,i1 eem,i2 eem,i3 eem,i4 eem,v1 eem,v2
T
eem,T 1 eem,t1 eem,t2 eem,t3 eem,t4
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Aem fem in (A.1) is just a matrix representation of the connections
between the faults and the effects described in the FMEA table presented in this section.
The propagation matrix Aem dy is defined below,
T
0 1 0 0 0 0 0
1 0 0 0 0 0 0
Aem =
0
dy 0 1 0 0 0 0
0 0 1 0 0 0 0
To cover the disturbances affecting this component, the fault vector is extended with
a set of disturbing events. These are,
191
Chapter A: FMEA Tables Describing Faults in the System
dem2 Increased supply Changes in the supply grid. eem,i1 Increased current.
voltage. Decreased grid load
combined with too small grid
for the load.
de2 Decreased supply voltage. Changes in the supply grid. eem,i2 Decreased current.
Increased grid load combined
with too small grid for the
load.
Including these disturbing events in the fault vector, it becomes,
T
fem = fem 0 T
dem1 dem2
Likewise the matrix Aem
fem is extended with rows according to the above table.
Load torque
Mechanical
Motor speed
dynamics
Motor torque
Figure A.2: The input/output structure of the part modelling the dynamics of the rotating
parts of the pump.
this component is included in the analysis to cover the signal structure of the system,
when building the functional model of the centrifugal pump. This means that no faults
are identified in the component.
The propagation matrix representation of the component is given by the following
logical expression,
dy dy
edy Ady dy dy
em eem + Amm emm + Ash esh + Ai ei + Amp emp (A.2)
192
Section A.3: Mechanical part of the motor
0 0 0 0 0 0 0 0 0 0 1
0 1 0 0 0 0 0
1 0 0 0 0 0 0
Ady =
0
mm 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0
Ady =
0
sh 0 0 0 0 1 0 1 0
0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 1 1 0 0 0 0
0 0 0 0 0 1 0 0 0 0 0 0
Ady
=
i 0 0 0 0 0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0
0 1 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0
Ady =
0
mp 0 1 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0
193
Chapter A: FMEA Tables Describing Faults in the System
Mechanical vibrations
Motor speed Mechanical part
Bearing temperature
of the motor
Load torque
Figure A.3: The input/output structure of the mechanical part of the induction motor.
The propagation matrix representation of the component is given by the following logi-
cal expression,
emm Amm mm
fmm fmm + Ady edy (A.3)
where the fault vector fmm and the effect vector emm equals,
T
fmm = fmm1 fmm2
T
emm = emm,t1 emm,t2 emm,t3 emm,t4 emm,T 1 emm,v1 emm,v2 .
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Ammfmm in (A.3) is just a matrix representation of the connections
between the faults and the effects described in the FMEA table presented in this section.
The propagation matrix Amm dy is defined below,
T
1 0 0 0 0 0 0
0 1 0 0 0 0 0
Amm
dy = 0 0 1 0 0 1 0
0 0 0 1 0 0 1
194
Section A.4: The shaft mechanics
Mechanical vibrations
Motor speed
Shaft mechanics Load torque
Impeller speed
Shaft Mechanics
Fault Modes Fault Causes Possible Effects
fsh2 Misalignment between Overload due to sudden esh,t1 1st harmonic torque
the motor and pump. blocked rotor. oscillations.
Fault introduced during esh,v1 Increased 1st
Parts
fsh3 Bend shaft. Overload, due to for esh,t1 1st harmonic torque
example blocked rotation. oscillations.
esh,v1 Increased 1st
harmonic vibrations.
The propagation matrix representation of the component is given by the following logi-
cal expression,
esh Ash sh
fsh fsh + Ady edy (A.4)
where the fault vector fsh and the effect vector esh equals,
T
fmm = fsh1 fsh2 fsh3
T
emm = esh,s1 esh,s2 esh,s3 esh,s4 esh,s5 emm,t1 emm,v1
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Ash
fsh in (A.4) is just a matrix representation of the connections
195
Chapter A: FMEA Tables Describing Faults in the System
between faults and the effects described in the FMEA table presented in this section.
The propagation matrix Ashdy is defined below,
T
1 0 0 0 0 0 0
0 1 0 0 0 0 0
Amm =
0
dy 0 1 0 0 0 0
0 0 0 1 0 0 0
Figure A.5: The input/output structure of the part describing the hydraulics of the cen-
trifugal pump.
fi2 Impurities fixed on the Impurities in the water. ei,h4 Harmonic pressure
impeller, causing inbalance. oscillations.
ei,t4 1st harmonic torque
oscillations.
ei,t5 higher harmonic torque
oscillations.
ei,v1 Increased 1st harmonic
vibrations.
ei,v2 Increased higher
harmonic vibrations.
196
Section A.5: Hydraulics of the Centrifugal Pump
fi3 Wear of the impeller. Sand and other impurities in ei,h2 Decreased pressure.
the water. ei,h4 Harmonic pressure
oscillations.
ei,t2 Decreased torque load
from the impeller.
ei,t4 1st harmonic torque
oscillations.
ei,t5 higher harmonic torque
oscillations.
ei,v1 Increased 1st harmonic
vibrations.
ei,v2 Increased higher
harmonic vibrations.
fi4 Blocked or partial Larges obstical in the liquid. ei,h2 Decreased pressure.
blocked flow field inside the ei,h4 Harmonic pressure
impeller.
oscillations.
ei,t2 Decreased torque load
from the impeller.
Parts
fi5 Blocked impeller Sand and other impurities in ei,h3 Zero pressure.
rotation. the water.
ei,t1 Increased torque load
from the impeller.
fi6 Wear of the sealing ring. Long running time / normal ei,h2 Decreased pressure.
wear. ei,t2 Decreased torque load
Overheated, due to for from the impeller.
example dry running.
197
Chapter A: FMEA Tables Describing Faults in the System
The propagation matrix representation of the component is given by the following logi-
cal expression,
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Aifi in (A.5) is just a matrix representation of the connections be-
tween the faults and the effects described in the FMEA table presented in this section.
The propagation matrices Aish and Aiip are defined below,
T
1 0 0 0 0 1 0 0 0 0 0 0
0 1 0 0 0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0 1 0 1 0
Aish =
0 0 0 1 0 0 0 0 0 1 0 1
0 0 1 0 0 0 0 1 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
T
0 0 1 0 0 0 0 1 0 0 0 0
0 1 0 1 0 0 1 0 0 0 0 1
0 0 0 0 1 0 0 0 0 0 0 0
Aiip =
0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0
To cover the disturbances affecting this component, the fault vector is extended with
a set of disturbing events. These are,
198
Section A.6: Mechanical Part of the Pump
di1 Decreased flow through Changes in the hydraulic ei,h1 Increased pressure.
the pump. system in which the pump is ei,t2 Decreased torque load
placed. from the impeller.
di2 Increased flow through Changes in the hydraulic ei,h2 Decreased pressure.
the pump. system in which the pump is
ei,t1 Increased torque load
placed. from the impeller.
Likewise the matrix Aifi is extended with rows according to the above table.
Figure A.6: The input/output structure of the part including all mechanical components
in the pump, not directly involved in the pressure generation.
fmp2 Inlet flow equal to zero. Error in the hydraulic emp,tp1 Increased bearing
system in which the pump is temperature.
placed.
199
Chapter A: FMEA Tables Describing Faults in the System
fmp3 Ware of the bearings in Sand in the liquid. emp,t3 1st harmonic torque
the pump. Wear out due to long runing oscillations.
time. emp,t4 higher harmonic
torque oscillations.
emp,tp1 Increased bearing
temperature.
emp,v1 Increased 1st
harmonic vibrations.
Parts
fmp4 Ware of seals. Over heated seals, due to for emp,q1 Leakage flow from
example dry running or zero the pump.
flow.
Wear out due to long runing
time.
fmp5 Rub impact between Dirte fixed between the emp,t1 Increased torque load
the impeller and the cassing. rotating and stationary parts of from the mechanical parts of
the pump. the pump.
Bend shaft due to overload, emp,t3 1st harmonic torque
cause by for example large oscillations.
obsticals in the liquid. emp,v1 Increased 1st
harmonic vibrations.
The propagation matrix representation of the component is given by the following logi-
cal expression,
emp Amp mp mp
fmp fmp + Ash esh + Ai ei (A.6)
where the fault vector fmp and the effect vector emp equals,
T
fmp = fmp1 fmp2 fmp3 fmp4 fmp5
T
emp = emp,t1 emp,t2 emp,t3 emp,t4 emp,T 5 emp,v1 emp,v2 emp,q1
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Amp fmp in (A.6) is just a matrix representation of the connections
between the faults and the effects described in the FMEA table presented in this section.
200
Section A.7: Inlet Part of the Pump
T
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 1 1 0
0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0
Amp
i =
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
Inlet flow
Impeller eye pressure
Inlet part of the
Inlet pressure pump
Pressure drop at inlet
Figure A.7: The input/output structure of the part including the component forming the
inlet of the pump.
201
Chapter A: FMEA Tables Describing Faults in the System
fip2 Low pressure at the inlet Poor dimensioned system. eip,e2 To low impeller eye
of the pump. pressure.
Low water level in the inlet
tank or well.
Parts
fip3 Opstruction of the inlet Clogging in the inlet part, eip,e2 To low impeller eye
of the pump. due to impurities in the water. pressure.
eip,h2 The pressure drop at
the inlet part is increased.
The propagation matrix representation of the component is given by the following logi-
cal expression,
eip Aip
fip fip (A.7)
where the fault vector fip and the effect vector eip equals,
0
T
fip = fip1 fip2 fip3
T
eip = eip,e1 eip,e2 eip,e3 eip,h1 eip,h2 eip,h3
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Aip fip in (A.7) is just a matrix representation of the connections
between the faults and the effects described in the FMEA table presented in this section.
To cover the disturbances affecting this component, the fault vector is extended with
a set of disturbing events. These are,
Disturbing events Causes Effects
dip1 Decreased flow through Changes in the hydraulic eip,h3 The pressure drop at
the pump. system in which the pump is the inlet part is decreased.
placed.
dip2 Increased flow through Changes in the hydraulic eip,h2 The pressure drop at
the pump. system in which the pump is the inlet part is increased.
placed.
202
Section A.8: Outlet part of the Pump
dip3 High frequency pressure High frequency pressure eip,e3 High frequency
oscillations. noise, due to for example pressure oscillations at the
cavitation in valves. impeller eye.
Leakage flow
Pressure drop at outlet
Outlet part of the
Inlet flow pump
Outlet flow
Figure A.8: The input/output structure of the part including the component forming the
outlet of the pump.
fop2 Leakage on the outlet Waer of the outlet part. eop,q2 The outlet flow is
pipe. descreased.
Parts
fop3 Opstruction of the outlet Clogging in the outlet part, eop,h2 The pressure drop at
pipe. due to impurities in the warter. the outlet part is increased.
203
Chapter A: FMEA Tables Describing Faults in the System
The propagation matrix representation of the component is given by the following logi-
cal expression,
eop Aop op
fop fop + Amp emp (A.8)
where the fault vector fop and the effect vector eop equals,
0
T
fop = fop1 fop2 fop3
T
eop = eop,h1 eop,h2 eop,h3 eop,q1 eop,q2
A description of each fault and effect is found in the above FMEA table. The fault
propargation matrix Aop fop in (A.8) is just a matrix representation of the connections
between the faults and the effects described in the FMEA table presented in this section.
The propagations matrix Aop mp is defined below,
T
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
Aop
mp =
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 1
To cover the disturbances affecting this component, the fault vector is extended with
a set of disturbing events. These are,
dip1 Decreased flow through Changes in the hydraulic eop,h3 The pressure drop at
the pump. system in which the pump is the outlet part is decreased.
placed.
dip2 Increased flow through Changes in the hydraulic eop,h2 The pressure drop at
the pump. system in which the pump is the outlet part is increased.
placed.
204
Section A.9: Difference pressure
Impeller pressure
Inlet pressure drop Outlet part of the
Difference pressure
Outlet pressure drop pump
Figure A.9: The input/output structure of the part modelling the dynamics of the rotating
parts of the pump.
The input/output structure of this part is shown in Fig. A.9. As described in chapter
4 this component is included in the analysis to cover the signal structure of the system
when building the functional model of the centrifugal pump. This means that no faults
are identified in the component.
edy = Adh dh dh
i ei + Aip eip + Aop eop (A.9)
T
edh = edh,h1 edh,h2 edh,h3 edh,h4 edh,h5 edh,h6
205
Chapter A: FMEA Tables Describing Faults in the System
T
1 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 0
Adh
i =
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
T
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
Adh
ip =
0 0 0 0 0 1
0 1 0 0 0 0
1 0 0 0 0 0
T
0 0 0 0 0 1
0 1 0 0 0 0
Adh = 1 0 0 0 0 0
op
0 0 0 0 0 0
0 0 0 0 0 0
206
Appendix B
Mathematical Tools
This appendix contains a short presentation of two mathematical tools used in this thesis.
These are a CUSUM alogithm, and some notes on Linear Matrix Inequalities (LMI).
H0 : = 0
H1 : = 1
These hypotheses can be tested using a recursive algorithm called the CUSUM algo-
rithm. One from of this algorithm is given by,
gk1 + ln p1 (yk ) if gk1 + ln p1 (yk ) > 0
p0 (yk ) p0 (yk )
gk = p (y ) (B.1)
0 if gk1 + ln p1 (ykk ) 0
0
207
Chapter B: Mathematical Tools
1 (y)2
p (y) = e 22
2
where is the mean value of the residual. Therefore it can be deduced that = 0 = 0
in the no fault case, and = 1 6= 0 in the faulty case. It is assumed that the variances
of the residual signals are known and do not change due to faults, i.e. in the above
expression is constant and known. When these properties are fulfilled the log-likelihood
in the CUSUM algorithm (B.1) is given by,
The signal gk in the CUSUM algorithm (B.1) is bounded from below by 0. In the
algorithm used in this work the signal gk is also bounded from above by h, resulting in
the following algorithm,
0 if gk1 12 yk 21 0
1 1
gk = h if gk1 2 yk 2 h (B.2)
1 1
gk1 2 yk 2 otherwise
This algorithm is a discrete algorithm. However, only continuous systems are considered
in this work. Therefore, for
the sake of cosistency a continuous version of (B.2) is given
below. Here s(t) = 12 y(t) 21 is used.
dg 0 if (g(t) = 0 s(t) 0) (g(t) = h 0 s(t))
= (B.3)
dt s(t) otherwise
208
Section B.2: Linear Matrix Inequalities
where
x = x1 x2 xm is a vector of real numbers.
F0 , , Fm are real symmetric matrices, i.e. Fi = FTi Rnn , i = 1, , m
for some n Z+ .
The inequality 0 in (B.4) means positive definite, i.e. uT F(x)u < 0 for all
u Rn . Equivalently, the largest eigenvalue of F(x) is negative.
This is stated in slightly more general in the following definition.
Definition B.2.1 (Linear Matrix Inequality) (Scherer and Weiland, 1999, Sec. 1.3) A
Linear Matrix Inequality (LMI) is an inequality,
F(X) 0 (B.5)
where F is an affine function mapping a finit dimensional vector space V to the set
S n = {M | M = MT Rnn }, n > 0 of real symmetric matrices.
The feasibility of the LMI defined in Definition B.2.1 denotes the existence of at least
one X which fulfills (B.5). Likewise the LMI is infeasible if such an X does not exist.
If an affine performance function on the form c(X) is defined, algorithms exists which
can solve the minimization problem minX c(X) constrainted by (B.5).
LMIs as defined in Definition B.2.1 can be used for checking stability of linear
systems with unknown and variating parameters. Such a system is shown below,
x = A((t))x (B.6)
where the state matrix A((t)) is a function of the real valued parameter vector . In all
physical system this parameter vector belongs to a bounded set , i.e. .
Suppose that the state matrix A((t)) is an affine function of the parameters, i.e.
Pk
A((t)) = A0 + j=1 j (t)Aj for all , then (B.6) is refered to as an affine
parameter dependent model. Suppose that the unknown parameters j j = 1, , k in
this expression are bounded on an interval, i.e. j [ j , j ]. If the unknown parameters
are bounded on this interval the corners of the uncertainty region is given by the set,
0 = = (1 , , k ) | j { j , j } (B.7)
Then the uncertainty region = co{0 }, where co denotes the convex hull of the
generators 0 (Scherer and Weiland, 1999).
The stability of (B.6), when A((t)) is an affine parameter dependent model, is
formulated by an LMI in the following proposition.
209
Chapter B: Mathematical Tools
Proposition B.2.1 (Scherer and Weiland, 1999, Sec. 2.4) if (B.6) is an affine parameter
dependent model then it is quadratically stable if and only of there exists P 0 such
that
A()T P + PA() 0
for all 0 , where 0 contains the generators of the convex hull, i.e. = co{0 }.
This proposition is motivated by the Lyapunov stability criterion. When Proposition
B.2.1 holds, system (B.6) is stable for arbitrary fast variations of . However, if the
parameters are unknown and bounded on a set, but not variating Proposition B.2.1 is to
restricted, i.e. system (B.6) can be stable for all even though Proprosition B.2.1
does not hold.
The following proposition states the stability conditions for (B.6) if the parameter
vector is assumed constant, i.e. = 0.
Proposition B.2.2 (Scherer and Weiland, 1999, Page 62) If (B.6) is an affine parameter
dependent model and Rk is the uncertainty set given by = co{0 }, where 0
is defined in (B.7), then system (B.6) is affinely quadralically stable if there exist real
matrices P0 , , Pk such that,
0 = { = (1 , , k ) | i [i , i ]} (B.9)
Then the uncertainty region = co{0 } is the set of possible variation rates of the
parameters . When the variation rates are limited on this set the following proposition
states the stability conditions for system (B.6).
Proposition B.2.3 (Scherer and Weiland, 1999, Page 64) If (B.6) is an affine parameter
dependent model and Rk is the uncertainty set given by = co{0 }, where 0
210
Section B.2: Linear Matrix Inequalities
is defined in (B.7), then system (B.6) is affinely quadralically stable if there exist real
matrices P0 , , Pk such that,
211