Faults Detection and Isolation computational tool using Neural Networks and State
Gloria Mousalli-Kayat* Jesús Calderón-Vielma**
Francklin Rivas-Echeverría*** Addison Ríos-Bolívar ***
Universidad de Los Andes
*Departamento de Medición y Evaluación
**Departamento de Circuitos y Medidas
***Departamento de Sistemas de Control
Mérida, Venezuela 5101
Abstract:-This work presents the design of a computational tool for fault Detection and Diagnostic in industrial
processes, through the integration of a virtual instrument developed in LabVIEW™ and a computer application in
MATLAB® for simulating an industrial process. Additionally, two faults detecting filters were developed using
MatLab®: one based on state observers and another based on a heuristic method and implemented through a
neural network. This tool has been made in English and Spanish in order to be handle by both language users.
Key words:- LabVIEW™, MatLab®, Integration, Fault Detection, State Observers, Neural Networks.
1 Introduction faults detection and diagnostic using virtual system
and Section 4 contains the corresponding
One type of programs that have had great growing in conclusions.
the last decades are the simulators, which try to
support the learning, emulating reality situations . 2 Faults detection
These programs have wide applications in the
engineering field, because they can be used for It is well-known that the operational reliability
modelling and simulating different kind of processes should be conformed by: the correct operation of the
using a personal computer (PC). processes, the appropriate control systems and the
coordination. This whole infrastructure is held by
Simulators allow to create similar atmospheres as the diverse support systems inside an integral automation
found in industrial control rooms and using them, it structure, where the information and its exchanges
is feasible to accumulate knowledge and experiences are considered outstanding, from the point of view of
that can be used in real conditions. reliability, security and productivity. In any level of
production chain, this information should be
In this work the capabilities of LabVIEW™ are managed for maintaining high efficiency indexes and
used for developing human-machine interfaces operational productivity.
(HMI), combined with a processes simulation
program developed using MatLab® for creating a Inside a mark of reliable and safe operation, there
computational tool, with the main objective of should be presented the systems that allow the events
helping the user with the fault detection and recognition, which should guide the decision-
diagnosis methods. The assistance is achieved makings when the behavior of the productive process
through a group of tests over a process and the user is affected by the presence of any adverse
will observe the behavior of the system under eventuality. Since the reliability is very near to the
different fault conditions. concept of security, then, it is fundamental to provide
industrial processes with demanding mechanisms of
The paper is structured as follows: Section 2 presents security whose basic elements are the Supervision,
a brief introduction to fault Detection. In section 3 it Diagnostic and Detection (SDD) systems; use the
is described the developed computational tool for indicators and the measured variables of the
processes, for maintaining a continuous and constant these techniques are in the building phase or for
supervision of the evolutionary behavior in the having very accurate models.
production time, in order to report any behavior that
is considered abnormal. In analytic techniques, all the information coming
from the processes measurement devices is used for
The SDD systems are based in their capacity to obtaining a mathematical model for diagnostic. The
respond under unexpected situations concerning the processes used for generating these residuals are [9,
process behavior, so their main task is the Faults 10, 11]:
Diagnosis and Detection, (FDD). A FDD system, as
the one shown in the Figure 1, uses the 1. The direct substitution in the model equations.
measurements of the process in order to produce
residuals, which, by means of evaluation functions 2. The use of the model together with the real
and decision logics, looks for the faults process, so that in both the same inputs are applied.
identificability and separability. So, any system that
allows, starting from measured variables of the 3. The use of a state observer. This is an extension of
processes, to generate residuals and to evaluate this the method of the parallel model. The main idea is
residuals deeply, considering decision makings for the residual generation, with precise directional
faults recognition is denominated Faults Detection properties, by means of an appropriate selection of
and Diagnostic Filter. the observer's gain.
4. The conception of an inverse model in order to
reconstruct the faults.
2.1 Observers-based Filter design
The state observers are analytic techniques based-on
the fault detection and diagnostic (FDD) filters
design. The FDD filters design can be divided in two
stages: the first phase is the residuals generation
(detection problem). The second stage is the
residuals evaluation in order to determinate the
Figure 1. SDD System origin of the faults, (faults separation problem). So,
the residuals are scalar or vector signals that contain
From the point of view of comparison-based the information about the time and localization of the
residual generation, the filters design techniques for faults. In principle, the residuals should be zero in
FDD can be classified as: absence of faults and, obviously, different from zeros
when some fault appear .
1.- Methods based on Physical
Redundancy: In this methods, it is used several Under those premises, the state observers can be
physical sensors of the devices and systems under used for residuals generation. The idea is to build a
study. These residuals are obtained by the complete order classic observer, for the system given
comparison between the answers of the different in (1), using the output variables y(t) and the control
elements. These techniques have the main variables u(t), in order to produce a vector of
inconvenience of high costs involved in their estimated states. The residuals are obtained
implementation and pursuit. comparing the estimated output with the measured
output of the physical plant.
2.- Methods based in Models: In this x(t ) Ax (t ) Bx (t ), x(0) x0
methods, it is produced estimated values for the y(t) Cy(t) (1)
processes variables and are used for the residuals
generation, by means of their comparison with the
Then, for the system (1) a gain matrix D nxq
measured outputs. The main inconveniences of
exists in such a way that the estimation x (t ) of the
state vector x(t) will be the solution for complete Fault Y1 Y2 Yi Yn
order observer's equation:
ˆ None -1 -1 -1 -1
x(t ) Ax(t ) Bx (t ) D ( y (t ) Cx(t )),
ˆ ˆ (2) Fault 1 1 -1 -1 -1
Fault 2 -1 1 -1 -1
The outputs of system (2) are the estimated outputs
and the observer gain matrix D should be selected
appropriately. So, defining the error signal by: Fault i -1 -1 1 -1
e(t ) x(t ) x(t )
which produces an innovation in the output defined Fault n -1 -1 -1 1
e(t) (A DC)e(t) A f f p B f f a DC f f s
Table 1. Neural Networks design for Fault detection
η(t) Ce(t) C f f s (4)
then, the error dynamics and the corresponding y1
output error will be given as S
η(t) y(t) y(t)
If D is selected in such a way that (A-DC) is stable,
that means, all their eigenvalues has negative real
part. In the limit, t, the estimation error will be xn yn
null (e(t)=0). In this case, it is said that the observer S
is exponential or asymptotic. Since for t <t0, the
process has a normal operation, that is, faults don't
exist; in that moment the residuals or the innovation Figure 2. Fault detection Scheme using Neural
of the output is approximately zero. When any fault Networks
appears, in t t0, the residual is different from zero
and it is favorable for faults detection .
3 Designed Computational Tool
2.2 Knowledge-based Methods
Knowledge-based methods are useful in those cases
where is difficult to find an analytic model for A computational tool was developed for using
obtaining the residuals. These methods are based on process faults detection and diagnostic methods.
the existing knowledge concerning the behavior of The presented computational tool allows the user to
the system outputs, this knowledge will allow to interact with the process through the Human
infer the process operational conditions. One of the Machine Interface (HMI) designed using
techniques that can be employed for this is Artificial LabVIEW™. The process is simulated in MatLab®.
Neural Networks [1,13]. The most used approach
for fault detection consists of a neural network that The main panel of the computational tool is
possesses as inputs the variables of the process used depicted in Figure 3. This screen presents an
for the detection and as outputs a groups of signals introduction to fault detection and diagnostic
according to the system operational condition. On methods. The user will be able to learn the different
the other hand, each desired output should be techniques using the connections to the based-on
chosen in such way that represents a particular redundancy and heuristic methods information.
condition of the process. For example, 1 represent From this main screen the user will have access to
the presence of a fault and –1 its absence, as can be the process that was chosen for implanting the faults
seen in next Table. systems. All the information in the computational
tool has been made in English and Spanish, in order
to increase the number of users that can use it.
The possible faults that have been considered in this
system are the blockage of any of the four valves
individually or at the same time.
Figure 3. Computational Tool Main Panel The equations that represent the system of Figure 4
The application example presented in this work 1 1 1 1 1
consists on a system with three interconnected
h1 (t ) ( )( )h1 (t ) ( )h2 (t ) ( )u (t )
r0 r1 a1 r0 a1 a1
tanks, as is depicted the Figure 4 .
1 1 1 1
h2 (t ) ( )h1 (t ) ( )( )h2 (t )
r0 a2 r0 r2 a2
1 1 1
h3 (t ) ( )h1 (t ) ( )h2 (t ) ( )h3 (t )
r1a3 r2 a3 r3a3
Under normal operational conditions, the four
valves that compose the system are open and it is
assumed that the supply of liquid is constant. As
the time continues, the levels in the three tanks are
increased, and the supply of liquid from tank 1 to
tank 2 neither the supply from tanks 1 and 2 to tank
3 are interrupted.
Figure 4. Three interconnected Tanks System
The system presented in Figure 4 consists of three Figure 5 illustrates the HMI designed for this
tanks, where the tank 1 receive a constant flow u(t) = process using LabVIEW. The user will be able to
5000 cm3/s, and it feeds tank 2 and tank 3. It is generate any of the four possible faults and to
assumed that the tanks have the following observe the behavior of the tanks levels.
dimensions: Additionally, the user will have a help file that
includes the system equations and the tank levels
Tank Area Height Initial behaviors under different faults conditions for the
1 2500 cm2 100 cm 0.5 cm
2 2000 cm2 80 cm 4.5 cm
3 3000 cm2 80 cm 0.2 cm
Starting from these values and considering that the
tank 1 flow is equal to 60% of the input flow u(t),
and the tank 2 flow are equal to 40% of the input
flow u(t); it was obtained the resistances values and
flows of each one of the four valves.
 Aguilar J., Rivas F, Introducción a las técnicas de
Computación Inteligente. Meritec, Venezuela
 Carbonell, J.R, AI in CAI: An Artificial Intelligence
Approach to Computer-Assisted-Instruction. IEEE
Transactions on Man-Machine Systems, 190-202
 Galvis, A., Ingeniería de Software Educativo.
Ediciones Uniandes, Colombia p 22-23 (1992).
 National Instruments Corporation, LabVIEW. User
Manual. USA (2000).
 The MathWorks Inc., Matlab Language of Technical
Computing. USA (1999).
Figure 5. Virtual Instrument developed using  Petzold, C., Programación en Windows® 95,
LabVIEW MacGraw-Hill Interamericana de España, S.A,
The computational tool implements the designed  Calderón-Vielma, J. Viloria et al, Integración de
detecting filters using the two methods introduced herramientas de programación para la enseñanza
previously. One of the filters is based on analytic de procesos. XVIII Interamerican Congress of
redundancy methods that calculate the residuals Chemical Engineering, Puerto Rico (1998).
starting from a state observer. The other filter is  Calderón-Vielma, J, Laboratorio Virtual para la
based on a heuristic method and implemented enseñanza de Automatización e Instrumentación
Industrial. Coloquio de Automatización y Control,
through a neural network; this neural network was Venezuela (1999).
trained using different operational conditions. Both  Rios, A., Sur la Synthèse de Filtres dde Détection de
filters are simulated in MatLab® and when those Défaillances. Ph.D Thesis, Université Paul
filters detect a fault condition it sends a signal to Sabatier, Toulouse. France. (2001).
LabVIEW™ and immediately it turns on a light  Szigeti F., J. Bokor, Edelmayer A. And Tarantino
signal indicating the fault that has happened. Using R., Fault detection Filter Design for Linear Time
this system, the users can evaluate the filters Varying Systems: Algebraic-geometric. 4th
operation and efficiency based on the obtained European Control Conference, Belgium (1997).
results.  Keviczky L., Bokor J., Edelmeyer A. and Szigeti F,
Detection Filter Design in System Perturbation
Application to Robust Change Detection and
4 Conclusions Identification, Proc. 12th IFAC World Congress
Vol 7 517-520, Australia (1992).
The designed computational tool for fault detection  Massoumnia M. A., A geometric approach to the
and diagnostic can be used for personal training in synthesis of failure detection filters. IEEE Trans.
this area, allowing on-line interaction with Aut. Control. AC-31.9. 839-846 (1986).
 Rios, A., Mousalli G. and Rivas F., Invertibility and
MatLab®. Neural Networks based FDI Filter. IASTED-ISC,
The use of a computer-based tool, allows an  Mousalli G., Calderón J., Ríos A., Rivas F.,
interactive way for studying fault detection and Learning Faults Detection and Diagnostic using
diagnostic methods based on state observers and Virtual System. Accepted for Publication on IJEE
artificial neural networks. Special Issue on Virtual Instrumentation. USA
This type of computational tools helps the users to  Novoa D., Pérez A., Rivas F., Fault detection
interact with industrial processes, to modify its scheme using Neo-fuzzy neurons. Proceedings of
operational conditions and to verify the obtained IASTED International Conference on Intelligent
Systems and Control. Hawaii USA (2000).
faults and its consequences. This helps to obtain an
inspection-based learning of real systems