Brazilian Journal

of Chemical ISSN 0104-6632

Engineering Printed in Brazil
www.abeq.org.br/bjche

Vol. 25, No. 04, pp. 777 - 788, October - December, 2008

DESIGN OF A MULTI-MODEL OBSERVER-BASED

ESTIMATOR FOR FAULT DETECTION AND
ISOLATION (FDI) STRATEGY: APPLICATION TO
A CHEMICAL REACTOR
Y. Chetouani*
Université de Rouen, Phone: (0033)23528466, Fax: (0033)235147191,
Rue Lavoisier 76821, Mont Saint Aignan Cedex, France.
E-mail: Yahya.Chetouani@univ-rouen.fr

(Received: February 1, 2007 ; Accepted : March 5, 2008)

Abstract - This study presents a FDI strategy for nonlinear dynamic systems. It shows a methodology of
tackling the fault detection and isolation issue by combining a technique based on the residuals signal and a
technique using the multiple Kalman filters. The usefulness of this combination is the on-line implementation
of the set of models, which represents the normal mode and all dynamics of faults, if the statistical decision
threshold on the residuals exceeds a fixed value. In other cases, one Extended Kalman Filter (EKF) is enough
to estimate the process state. After describing the system architecture and the proposed FDI methodology, we
present a realistic application in order to show the technique's potential. An algorithm is described and applied
to a chemical process like a perfectly stirred chemical reactor functioning in a semi-batch mode. The chemical
reaction used is an oxido reduction one, the oxidation of sodium thiosulfate by hydrogen peroxide.
Keywords: Safety; Reliability; Risk assessment; FDI method; Model-based approach; Extended Kalman Filter
(EKF).

INTRODUCTION isolation of faults in engineering and industrial

systems is a critical factor for avoiding product
A chemical process can only be implemented for deterioration, major damage to machinery, loss of
industrial application if a complete study has been production, performance degradation, poor plant
carried out to guarantee its safety and the quality of economy, environmental pollution and damage to
its products. Before this industrial stage, phenomena human health or even loss of life.
occurring during the chemical process are FDI is an active area of research due to growing
characterized to make sure that this process is the demand for safety and reliability and increasing
most adapted. However, plants in the chemical and complexity of process plants. Many FDI techniques
biochemical industries are becoming larger and more have been proposed in the literature. These techniques
complex. The growing safety and environmental can in general be classified as model-based
demands are forcing industry to look for more approaches, data-driven approaches, logic-based or
powerful and new techniques for the detection of information flow graphs, hardware redundancy,
process faults. A fault is defined as an unexpected knowledge-based systems and analytical redundancy
change of the system functionality which may be techniques. We are interested in the concept of so-
related to a failure in a physical component or in a called model-based FDI with reference to observer
system sensor or actuator. The early detection and and filter-based techniques (Chetouani, 2006a,
*To whom correspondence should be addressed
778 Y. Chetouani

Chetouani, 2006b). Model-based techniques all use by an extended Kalman filter, by which the variance
mathematical models of the plant being monitored. A of the measurements and parameters can be optimally
wide variety of these approaches have been proposed considered. Nyberg et al. (2004) used an extended
to tackle the isolation problem (Patton et al., 2000; Kalman filter for the air-path of a turbo-charged diesel
Chen et al., 1999; Isermann et al., 1997). However, engine. The engine is equipped with exhaust gas
the conceptual realization of these models can vary recycling (EGR) and a variable nozzle turbine (VNT).
according to the following approaches; the fault The faults considered were air mass-flow sensor fault,
detection filter (Liberatore et al., 2006; Henry et al., intake-manifold pressure sensor fault, air-leakage, and
2005), the parity space (Isadi et al., 2005), parameter the EGR-valve stuck in closed position. They showed
identification (Fantuzzi et al., 2002; Zogg et al., that the diagnosis system was successfully evaluated
2006), state estimation (Biagiola et al., 2006) and in a real car driving on the road. Alessandri (2003)
nonlinear techniques (Edelmayer et al., 2004; Korbicz presented a model-based method to detect faults in
et al., 2004). The model-based approach based on a nonlinear systems by means of a bank of estimators,
Kalman filter has been used in different type of which provide estimates of parameters that describe
applications, like the on-line failure detection in actuator, plant, and sensor faults. These estimators
nuclear power plants (Tylee, 1983); the on-line fault perform according to a receding-horizon strategy and
in a chemical reactor (Chetouani, 2004), in DC motors are designed using models of the failures. The
to predict failures by placing an exponential attenuator performances obtained in the estimation of the fault
at the output end of the motor model to simulate aging parameters by the proposed neural estimators and by
failures, resulting in small prediction errors (Yang, the extended Kalman filters are compared by means
2002), or in the detection of faults in turnouts in the of simulations with an application to underwater
infrastructure elements of railway systems (García et robotics. Shi et al. (2003) presented a filter framework
al., 2003; Pedregal et al., 2004). Simani et al. (2006) to estimate two-dimensional left ventricular
used the EKF for the dynamic system identification deformation from spline-regularized MRI (magnetic
and model-based fault diagnosis of an industrial gas resonance imaging) phase contrast velocity fields that
turbine prototype. A model-based procedure are constrained by segmented endocardial and
exploiting analytical redundancy for the detection and epicardial contours. After the conversion to state
isolation of faults in a gas turbine process was space representation, the extended Kalman filtering
presented. The main point of their work consists of procedures were adopted to linearize the equations
exploiting system identification schemes in and to provide the joint estimates in an approximate
connection with observer and filter design procedures optimal sense. They concluded experimentally that it
for diagnostic purpose. Zhan et al. (2006) presented a is possible to adopt this biomechanical model-based
model-based technique for the detection and diagnosis multiframe estimation approach to achieve converged
of gear faults under varying load conditions using the estimates because of the periodic nature of the cardiac
gear motion residual signal. A noise-adaptive Kalman dynamics. Bhagwat et al. (2003) proposed a model-
filter-based Auto-Regressive (AR) model was fitted to based fault detection scheme that involves
the gear motion residual signals in the healthy state of decomposition of nonlinear transient systems into
the gear of interest. It consists of fitting an AR model multiple linear modeling regimes. Kalman filters and
to gear motion residual signals and takes advantage of open-loop observers were used for state estimation
the NAKF to decorrelate the signal to produce a white and residual generation based on the resulting linear
Gaussian sequence. Pedregal et al. (2006) developed a models. Li et al. (2003) presented an on-line
complex prediction model based on the algorithm estimation of stirred-tank microalgal photobioreactor
Kalman filter for vibration data within the state space cultures based on dissolved oxygen measurement. An
class applied to real data in the petrochemical extended Kalman filter (EKF) was applied to provide
industry. The core of the system is a model to forecast optimal estimates of photobioreactor states, based on a
the state of the machine using data provided by the dynamic process model in conjunction with on-line
condition monitoring system at each moment in time. dissolved oxygen measurement. Jang et al. (2000)
Feil et al. (2006) presented a monitoring system based proposed a model-based tracking algorithm that can
on Kalman filtering for process transitions. They used extract trajectory information of a target object by
a model-based state-estimation algorithm to detect the detecting and tracking a moving object from a
changes in the correlation among the state-variables. sequence of images. They applied the Kalman filter to
Brännbacka et al. (2004) developed an advanced predict motion information, which was used
model to track iron and slag levels in the blast furnace efficiently to reduce the search space in the matching
hearth. The liquid level estimation problem is tackled process.

Brazilian Journal of Chemical Engineering

 Design of a Multi-Model Observer-Based Estimator for Fault Detection and Isolation (FDI) Strategy 779

In this paper, the basic idea of the adopted FAULT ISOLATION STRATEGY BY USING
approach is to reconstruct the outputs of the system THE MULTIPLE KALMAN FILTERS
from the measurement using observers and using
the residuals for fault detection (Mehra et al., 1971; A FDI system should perform two tasks, namely
Franck, 1990; Chetouani, 2004). This study also fault detection and fault isolation. The purpose of the
shows a methodology for tackling the isolation fault detection module is to determine that a fault has
issue of fault causes by combining the last occurred in the system. The second task is devoted to
technique and the technique using the multiple locate the fault cause. To achieve this aim, all the
Kalman filters for a nonlinear dynamic system. The available information from the process should be
usefulness of this combination is the on-line collected and processed to detect any change from
implementation of all the models of fault dynamics nominal behavior of the process. Model-based fault
if the statistical decision threshold on the isolation relies on mathematical models of the plant
standardized innovation exceeds a fixed value. In to identify a discrepancy between the nominal plant
other cases, one extended Kalman filter is enough and the plant when a fault occurs (Isermann, 2005).
to estimate the process state in order to avoid the The detection of faults is usually accomplished by
high numerical integration of all models. After evaluating the residuals that are sensitive to the
describing the system architecture and the proposed fault’s occurrence in the process dynamics. In the
methodology of the FDI, we present a realistic case of a normal behavior, the residuals have
application in order to show the technique's properties that are known a priori. Hence, any
potential. The purpose is to develop and test the sudden or abrupt change in the value of any control
FDI method on real incident data, to detect the parameter of the process implies some deviation of
occurrence of change, to pinpoint the moment it those properties from the normal mode profile. Fig. 1
occurred and to locate the fault cause. The shows the fault isolation method based on the
experimental results demonstrate the robustness of analysis of the residuals’ evolution between the real
the FDI method. output and the output estimated by the extended
The structure of the paper is as follows. First, the Kalman filter. This filter is an effective tool for
description of the FI strategy by using the multiple stochastic estimation of the state from noisy
Kalman filters is presented. Then the FD strategy by measurements. Because of its relative simplicity and
using the standardized innovations is presented robust nature, the Kalman filter has been widely used
followed by the description of the proposed FDI to obtain estimates of the state variables in practice.
method. Then a comprehensive experimental The filter’s number is defined from the exhaustive
validation of the proposed method is presented. list of all the models of faults and the model of the
Finally, conclusions are drawn in the last section. normal behavior of the process.

Figure 1: Scheme of the FDI strategy

Brazilian Journal of Chemical Engineering Vol. 25, No. 04, pp. 777 - 788, October - December, 2008
780 Y. Chetouani

The multiple filters method allows establishing a where γ k = (y k − h k (xˆ −k )) is the innovation between
link between the detected anomalies during the the measured output and the estimated one.
process and a given situation of the process. It
consists of having the observations processed by a ƒ The a posteriori state covariance is:
set of EK filters (King et al., 1990). Each EKF
corresponds to an assumption of a particular fault.
{y1(t)…yn(t)} represent the models of faults. During Pk+ = (I − K k H k (xˆ k− ))Pk− (4)
normal functioning, the EK filters generate
innovations which must be low. They depend only ˆ
where H(x(t), t) is the Jacobian matrix of the
on the intrinsic errors and the random variations of partial derivative of h according to the state
the measurement equipment. In other cases where vector:
the process dynamics are not respected, these same
signals evolve differently according to the magnitude
∂ h i (x(t), t)
of the effect of the fault on the general state of the ˆ
Hij (x(t), t) = (5)
plant. Consequently, the aim of this method consists ∂ x j (t)
x(t) = x(t)
ˆ
of determining the nearest filter which generates the
lowest innovation. Under normal operating conditions, the
By supposing that the behavior of a nonlinear
process is described by a finished set of models innovation γ k = (y k − h k (xˆ −k )) is a white noise
indexed i = 1,...., N ; and by the input u (k ) and the sequence with a zero mean and a variance:
output z(k) for k = 1, 2,... , the dynamics of a
nonlinear process can be expressed by the following: ηk = (H k (xˆ −k )Pk− H k T (xˆ −k ) + R) (6)

x i (t) = fi (x i (t), u i (t), t) + w i (t) (1) By noting that the model i is the model which
represents the dynamics of the real system, its
probability can be defined on-line according to a
zi (k) = h i (x i (t k )) + vi (k) k = 1, 2, ...
hypothesis test. Unlike the standard Kalman filter,
this method has for input the set
where x ∈ ℜ n is the state vector, with an initial value Ii (k) = ( u i (0),..., u i (k − 1), zi (1),..., zi (k) ) and for
of x 0 , z is the measurement vector, w(t) is the
output the probability of each model. By applying
system noise representing modeling error and Bayes’ law, the probability of each model is:
unknown disturbances, and v(k) is the measurement
noise. Both system and measurement noises are p ( zi (k + 1/ Hi ), Ii (k), u i (k) ) pi (k)
assumed to be independent random white noises with pi (k + 1) = N
(7)
zero mean and covariance matrices, respectively Q
∑ p ( z j (k + 1/ H j ), I j (k), u j (k) ) p j (k)
and R. The functions f and h denote, respectively, a j=1
nonlinear function of the state and the output. The
extended Kalman filter is essentially a set of
where the density of the conditional probability
mathematical equations that aims at minimizing the
p(zi (k + 1/ H i ), Ii (k), u i (k)) is calculated according
estimated error covariance in the state estimator
(Hovland et al., 2005). This estimation by EKF is to the nature of the noise. Its distribution follows a
given by the following equations: normal law:

The Kalman gain is: p ( zi (k + 1/ Hi ), Ii (k), u i (k) ) =

(8)
 1 
Kk = Pk− H k T (xˆ −k )(H k (xˆ −k )Pk− H k T (xˆ −k ) + R)−1 (2) exp  − γ i (k + 1)T η−i 1 (k + 1) γ i (k + 1) 
 2 
ƒ The a posteriori state estimate is: (2π) m/2
( det(ηi (k + 1) ) 1/ 2

xˆ +k = xˆ −k + K k (y k − h k (xˆ −k )) (3) m represents the dimension of the measurements

vector.

Brazilian Journal of Chemical Engineering

 Design of a Multi-Model Observer-Based Estimator for Fault Detection and Isolation (FDI) Strategy 781

FAULT DETECTION STRATEGY BY USING if the latter is slow, the size will be large. Otherwise,
THE STANDARDIZED INNOVATIONS when the dynamics evolves rapidly, samples will be
taken quickly in order to enable the decision to be made
Analysis of the generated innovation is required for as early as possible. ηs refers to the true mean of the
fault detection. In the absence of faults, the innovation
sample. Under the hypothesis ( H ), ηˆ s has a gaussian
o
value is near zero. A fault causes the process to behave
differently from normal model prediction, and the filter, distribution with a zero mean and a covariance
which compares the model and the process, generates E(ηˆ sηˆ sT ) = I / N . However, beyond a given level of
large innovation. Faults are detected by deviation of the acceptance, such as above α (‘non-normal’ hypothesis
innovation from this normal value. In practice statistical H1 ), the ‘normal’ hypothesis ( H o ) is rejected:
analysis of innovation such as mean moving average
mean, whiteness ratio can also be used as the fault
detection signal. But in the case study (Chetouani, ηˆ s > α (12)
2004), it was observed that the innovations are
themselves a good indication of the process state The threshold α depends on the probability of
(normal or faulty). In this case, the statistical threshold false alarms. Threshold limits can be defined for
based on innovations allows delimiting two distinct innovations for a normal behavior. Threshold limit
regions; the first region is called the safe region where selection, in most cases, is dependent on the process
the innovation variation is considered to be acceptable. at hand (Himmelblau, 1978). The threshold
The second region is not acceptable (fault region) and hypothesis accounts for innovation changes due to
is where the innovation variation exceeds the statistical noise and some process-model mismatch. In other
threshold. Considering a fault region π in the sample words, threshold limits can be defined and based on
space ℜn , an acceptable region Cπ is defined as the minimum process deviations that are not acceptable.
complement of π in ℜn . Let M be a point of size n
THE PROPOSED FDI APPROACH
representing the sample. The sampling corresponds to a
real representation of the process. Notice that it is The standard multiple filters method is based on all
convenient to use the standardized innovation sequence the models (normal and faulty) of the mathematical
for the standardized hypothesis of statistical tests library of faults (or bank of filters). The on-line
(Chetouani, 2006a). This sequence is expressed as implementation of all the models is carried out upon
follows: starting the process. Another inconvenience of the EKF
compared to the linear filter is the high time required
ηs k = (H k (xˆ k− )Pk− H k T (xˆ −k ) + R) −1/ 2 for processing; the Kalman gain is related to the
(9) Jacobian matrices of the model based on the estimated
(y k − h k (xˆ k− )) = (ηk ) −1/ 2 γ k state. In fact, the extended Kalman filter is based on a
linearization, enabling the state estimation of the
with nonlinear system. Consequently, it cannot be
guaranteed that the estimation will converge. If the
E(ηs jηsTk ) = Iδ jk (10) filter converges, the precision of the estimation depends
on the system’s trajectory. In order to avoid this
numerical issue, this study proposes an algorithm
δ and I represent, respectively, the Kronecker
which scrutinizes the state of the standardized
symbol and the identity matrix. ηs k has a zero mean innovation i.e. the use of only one extended Kalman
and an unit variance. Any deviation from the normal filter. During a normal functioning, this filter allows
behavior causes a substantial change of these estimation of the measurable and non-measurable states
statistical properties. A test on the mean of the of the process (moles number, concentrations, or
normal law N(0, I) is carried out to verify whether others) and could be even used for the development of
the standardized innovation sequence (Chetouani, the control laws based on the process predictions. In
2006a) has a zero mean. The latter is estimated by: other cases when the statistical threshold for the
standardized innovation exceeds a fixed level of
1
N significance (15 %), including the measurement
ηˆ s =
N ∑ ηs
j=1
j (11) noises and the modelling errors, the fault isolation
module is engaged. The usefulness of this
combination is to avoid the matrix inversions and
where N is the sample size, which depends on the numerical integrations when the process is
system dynamics. The size of the representative functioning near to the normal mode. This
observation sample depends on the system response; combination is described by the diagram in Fig. 2.

Brazilian Journal of Chemical Engineering Vol. 25, No. 04, pp. 777 - 788, October - December, 2008
782 Y. Chetouani

Figure 2: The proposed FDI approach

APPLICATION OF THE PROPOSED FDI constant) can also be introduced. Supervision

APPROACH software also allows the regulation of the flow rate
introduction and the stirring rate of the reactor. It
Experimental Device: Chemical Reactor displays and stores data during the experiment for
further exploitation.
The chemical reactor represented in Fig. 3 works
under atmospheric pressure. It is a glass-jacketed Chemical Reaction Choice
reactor with a tangential input for heat transfer fluid.
The capacity of the reactor is two liters. It is In order to illustrate the proposed FDI approach,
equipped with an electrical calibration heating, a a chemical synthesis in a laboratory-size reactor
stirring system (anchor) and an input system (two was carried-out. The reaction chosen is a very
dosing pumps and two scales, with temperature exothermic oxido-reduction one (Aime, 1991), the
measurement at input). It is also equipped with Pt100 oxidation of sodium thiosulfate by hydrogen
temperature probes. Five probes are inserted through peroxide. This reaction can be expressed by the
the chemical reactor in the reaction mass, in the following equation:
cooling, in the condenser and in the flow rate
introduction. The heating-cooling system, which N a 2 S 2 O 3 + 2 H 2 O 2 → 0 .5 N a 2 S 3 O 6 +
uses a single heat transfer fluid, works within a (13)
temperature range from -15 to +200 °C. Supervision 0 .5 N a 2 S O 4 + 2 H 2 O
software allows the fitting of the parameters and
their instruction value. The parameters introduced in This equation is written as:
the supervision system are the control modes of the 2A + B → 0.5C + 0.5D + 2E , where the following
reaction mass temperature (isothermal or slope of the notations are adapted: A : hydrogen peroxide
temperature) and of the jacket fluid (constant or H 2 O2 ; B: sodium thiosulfate Na 2S2 O3 ; C :
slope of the temperature). The distillation mode (the
sodium persulfate Na 2S3O6 ; D : sodium sulfate
difference between the reaction mass temperature
and the fluid jacket temperature is maintained Na 2SO 4 ; E : water H 2 O .

Brazilian Journal of Chemical Engineering

 Design of a Multi-Model Observer-Based Estimator for Fault Detection and Isolation (FDI) Strategy 783

Figure 3: Diagram of the chemical reactor

Chemical Reactor Modeling

mR CpR
dTR (t)
dt
(
= −hR AR TR (t) − Tw (r , t) r =R1 )−
In order to guarantee that faults can be detected
−E

( n (M − χ)/ V)
and isolated, mathematical models of the process
under investigation are required, either in state space k0 e RT
o
α
( no (Z − χ))β V∆H (14)
or input-output form. State space descriptions
TF j
generally provide mathematically rigorous tools for
system modelling and residual generation that may
be used in fault detection of industrial systems, both
+Fj ∫ Cp dT + Kc( T − T (t))
TR
j a R

for noise-free measurements and noisy data

environment. The dynamic behavior of the chemical d Tf (t) •
reactor is modeled by a system of differential mf Cpf = mf Cpf ( Tfe (t) − Tf (t) ) +
equations translating molar and heat balances in the dt
(15)
reactor. The model used is based on the following
hypotheses (Chetouani, 2004): h f A f Tw ( r = R1 + e
− Tf (t) )
ƒ The reactor is perfectly stirred: reaction mass
temperature and concentrations are homogeneous ∂ 2 Tw (r, t) 1 ∂ Tw (r, t)
through the mass volume; + +
ƒ No phase change in the reaction mass; ∂r 2 r ∂r
(16)
ƒ The density, the specific heat of the cooling and 2
∂ Tw (r, t) 1 ∂ Tw (r, t)
the reaction mass heat are independent of the =
temperature. ∂z 2 aw ∂t
ƒ By taking into account the thermal inertia of the
wall, three energy balances (14, 15 and 16) were With boundary conditions in ℜ(O, r, z = 0) :
established, respectively, in the reaction mass, in the
jacket fluid and in the reactor wall. t = 0; Tw (r, 0) = (µ ln(r ) + τ) (17)
R1 ≤ r ≤ R1 + e

Brazilian Journal of Chemical Engineering Vol. 25, No. 04, pp. 777 - 788, October - December, 2008
784 Y. Chetouani

where significant if one is interested in fault detection and

isolation. Indeed, the tests are established by

µ=
( T (r, t)
w r = R1
− Tw (r, t) r =R
1+e
) (18)
analyzing the reconstruction error, which is sensitive
to the detected faults and also to the modeling errors.
R +e Before solving equations (14, 15 and 22) by the
ln  1  Runge-Kutta-Fehlberg method (Forsythe et al.,
 R1  1977), the equation system for the heat balance in the
reactor wall (16, 20 and 21) is transformed into an

τ=
(T (r,t)
w r =R1
ln(R1 + e) − Tw (r,t) r=R +e ln(R1)
1
) (19)
algebraic equation system. By employing the finite
difference method under its explicit formulation,
 R +e equations (16, 20 and 21) become, according to the
ln  1 
 R1  numerical scheme in Fig. 4:

t>0 ƒ Temperature of the reaction mass for an interior

nodal point ( R1 < r < R1 + e ):

(
hR AR TR (t) − Tw (r,t) r=R = −λw AR
1
) ∂ Tw (r,t)
∂ t r=R
(20)  ∆t a ∆t 
Tin +1 = 1 − 2a w 2 − w  Tin +
1
 ∆r r∆ r 
(23)
( ) ∂ T (r,t)
hf Af Tw (r,t) r=R +e − Tf (t) = −λwAf w
1 ∂ t r=R +e
(21)  ∆ t aw∆ t  n 
 a w 2 +
∆t  n
 Ti +1 +  a w 2  Ti −1
r∆ r 
1  ∆r  ∆r 

Using the normalized reaction extent ƒ where the subscript refers to the interior nodal
(Villermaux, 1993), the mass balance in the reaction point and the exponent refers to the time step. The
mass can be written according to an Arrhenius law: boundary temperature of the reactor wall at r = R1 is
given as follows:
dχ n 0
=
dt V
α
(
k 0 exp( −E / RTR ) M − χ ) ( Z − χ )
β
(22)
  1  
Tin+1 = 2Fo  BiR TRn + Tin−1 +  − BiR −1 Tin  (24)
Resolution of the Chemical System   2Fo  

The constituent order values for the reactants are ƒ The boundary temperature of the reactor wall at
σ = 1.5 and β = 0.6 when the concentration ratio r = R1 + e is given as follows:
between hydrogen peroxide and sodium thiosulfate is
maintained greater than 1.96 during the reaction
(Cohen et al., 1962). Under the same operating   1  
Tin+1 = 2Fo  Bif Tfn + Tin−1 +  − Bif −1 Tin  (25)
conditions as those of Cohen et al. (Cohen et al.,   2Fo  
1962) and by assuming that the rate coefficient is
governed by the Arrhenius law, Aime (Aime, 1991) where BiR = hRe/ λw and Bif = hf e/ λw represent the
has estimated the kinetic parameters (frequency
Biot numbers computed by using, respectively, the
coefficient k o = (7.25±0.5) × 1011 L/mol.s and
convective resistance of the reaction mass side
activation energy E = 79002±1254 J/mol), as well ( 1/ hR AwR ) and that of the cooling fluid side
as the thermodynamic reaction enthalpy ( ∆H = -
563±10 kJ/mol). The wall equation (16-21) is used in ( 1/ hf Awf ). Fo = a w∆ t / ∆ r2 is the Fourier number.
order to improve the model. In practice, the model of The convergence of the equation system (14, 15, 22,
a system provides only an approximation of the real 23, 24 and 25) depends on the stability and on the
behavior. Because the observer is built from this consistency of the equation sub-system (23, 24 and
model, the provided state and the output are sensitive 25). Consequently, ∆r and ∆t are taken in such a
to the errors coming from the difference between the way that the stability and convergence conditions are
system and its model. This point is particularly respected.

Brazilian Journal of Chemical Engineering

 Design of a Multi-Model Observer-Based Estimator for Fault Detection and Isolation (FDI) Strategy 785

R1+e
R1

r=0 r = R1 r = R1+e

Middle 1 Middle 2 Middle 3

Reaction mass Wall Cooling

Figure 4: Scheme of the numerical resolution

FDI Results the evolution of the innovation (Fig. 6). The latter is
processed by the EKF to detect an actual fault
In order to illustrate the pertinence of the condition, rejecting any false alarms caused by noise
proposed FDI approach, it is applied to detect and or spurious signals. The convergence of the
locate a fault due to a sudden change, which occurs innovation depends on the R / Q ratio. In fact, the
at 500 s, of the flow cooling during the phase of knowledge of these two variances constitutes a
preheating of a reaction mass. The normal conditions means to adjust the EKF. It represents an indication
are chosen corresponding to the flow Qvn = 0.40.10-3 of the filter’s degree of dependence on the
m3.s-1 and as non-normal conditions those which measurements; if R is small and Q is large, then
correspond to the flow Qvfault = 1.10.10-3 m3.s-1. information coming from the measurements is
Theoretically, this last flow, which provides the preponderant in the estimation, and if R is large and
measurement dynamics y s , is unknown by the Q small, the filter gives preference to the model
operator. The aim of the FDI method is to identify it rather than to measurements (Chetouani, 2006a).
by comparing its dynamics with those given in real- By analyzing the probability results (Fig. 7), it is
time by the mathematical library of different models. observed that the fault isolation phase starts at 820 s,
For this purpose, different dynamics corresponding i.e., 320 s after the start of the fault detection phase,
to different flows of the cooling are built. This flow in order to avoid numerical calculations and false
change does not modify the structure of the model alarms. In addition, it is noted that the fault dynamics
equations, only the partial transfer coefficient h f is ys = f (Qv fault ) is close to that corresponding to the
modified. The selected flows of the cooling are as flow ys = f (Qv1 ) because of the probability p(Qv1 ) ,
follows (Table 1). which is most important. The probability of the other
The results of different simulations are given as models decreases considerably according to their
follows (Fig. 5). respective variation compared to the output
The evolution of the probability is conditioned by ys = f (Qv fault ) , as shown in Fig. 7.
340
Qv 4 Qv n
Qv 3
ys
330 Qv 2
Temperature (K)

320

Qv 1
310

300

290
0 2000 4000 6000 8000 10000 12000 14000

Time (s)
Figure 5: Temperatures generated by the different dynamics

Brazilian Journal of Chemical Engineering Vol. 25, No. 04, pp. 777 - 788, October - December, 2008
786 Y. Chetouani

Table 1: Selected flows of the cooling

Volumic flow ( m3.s-1 ) 1.33.10-3 0.80.10-3 0.60.10-3 0.20.10-3

Dynamics Qv1 Qv2 Qv3 Qv4

0.20
Starting of the fault isolation phase (820 s)
0.10 Starting of the fault detection phase (500 s)

0.00
500 3000 5500 8000 10500
Innovations

Qv 1
-0.10 Time (s)
Qv 2

-0.20

Qv 3
-0.30

-0.40 Qv n

Qv 4
-0.50

Figure 6: Evolution of the generated innovations by the EK Filter

0.8

0.6
Starting of the fault isolation phase (820 s)
Probability

Starting of the fault detection phase (500 s)

0.4
p(Qv 2 )
p(Qv 1 )

0.2
p(Qv n )
p(Qv 3 )
p(Qv 4 )
0.0
500 3000 5500 8000 10500
Time (s)

Figure 7: Evolution of the different probabilities

CONCLUSION Kalman estimators and the standardized innovations

method. The proposed method allows detection of
In this paper, a model-based fault detection and the changes in the process behavior compared to the
isolation method for nonlinear systems has been normal one. The method also allows identification of
considered, based on the idea of generating residuals. the new regime describing most accurately as
The proposed method requires physical knowledge possible the evolution of the system. The analysis of
of the process under observation. First an accurate the probabilities shows that the FDI method adopted
model of the process dynamics has been constructed in this study allows considerable reduction in the
according to all the available information for the computing time. In fact, this method requests the
studied system. An extended Kalman filter is used library mathematics of faults 320 seconds after the
for state estimation and residual generation. Faults fault occurrence. This method can be applied to
are detected by a statistical threshold and isolated by many industrial processes (heat exchanger,
using a combination of two decision methods. This distillation column…) if a modeling of the process is
evaluation is accomplished by means of a bank of carried out. The FDI adopted approach is a model-

Brazilian Journal of Chemical Engineering

 Design of a Multi-Model Observer-Based Estimator for Fault Detection and Isolation (FDI) Strategy 787

based approach. However, in practical applications, χ Normalized reaction extent (-)

especially when considering large plants and ∆H Reaction enthalpy kJ.mol-1
complex systems, the straightforward application of ∆ variation (-)
such FDI techniques may be difficult. Thus, these
models are often very nonlinear and very complex,
and they can exploit hybrid structures to describe REFERENCES
accurately the behavior of the real-target system. A
viable and easy procedure for practical application of Aime, N., Aide à la conduite sûre des réacteurs
this technique is really necessary in complex chimiques discontinus en marche normale ou
applications. incidentelle vis-à-vis du danger d’emballement
thermique, Ph.D. Thesis, Université de
Technologie de Compiègne (1991).
ACKNOWLEDGEMENT Alessandri, A., Fault diagnosis for nonlinear systems
using a bank of neural estimators, Computers in
I acknowledge the critical comments of the Industry, 52, 271-289 (2003).
anonymous reviewers for their very good comments, Bhagwat, A., Srinivasan, R., Krishnaswamy, P. R.
which greatly improved the manuscript. I would like Multi-linear model-based fault detection during
to thank the edition staff for helping with the English process transitions, Chemical Engineering
of this manuscript. Science, 58, 1649-1670 (2003).
Biagiola, S., Solsona, J., State estimation in batch
processes using a nonlinear observer,
NOMENCLATURE Mathematical and Computer Modelling, 44,
1009-1024 (2006).
A Exchange area m2 Brännbacka, J., Saxén, H., Novel model for
3 -1 -1
ko frequency coefficient m .mol .s estimation of liquid levels in the blast furnace
E Energy activation J.mol-1 hearth, Chemical Engineering Science, 59, 3423-
Cp Specific heat J.kg-1.K-1 3432 (2004).
e Wall thickness m Chen, J., Patton, R.J., Robust model-based fault
F Flow feeding mol.s-1 diagnosis for dynamic systems, Kluwer
H Heat transfer coefficient W.m-2.K-1 Academic Publishers (1999).
Kc Conductance W.K-1 Chetouani, Y., Fault detection in a chemical reactor
N Mole number mol by using the standardized innovation, Process
M Molar mass kg.kmol-1 Safety and Environmental Protection, 84, 27-32
m Mass kg (2006a).
• Chetouani, Y., Application of the generalized
m Flow rate kg.s-1 likelihood ratio test for detecting changes in a
R Ideal gas constant J.mol-1.K-1 chemical reactor, Process Safety and
R1 Reactor radius m Environmental Protection, 84, 371-377 (2006b).
R Reaction rate mol.m-3.s-1 Chetouani, Y., Fault detection by using the
T Temperature K innovation signal: application to an exothermic
T Time s reaction, Chemical Engineering and Processing,
V Volume m3 43, 1579-1585 (2004).
Cohen, W. C., Spencer, R., Determination of
Subscript chemical kinetics by calorimetry, Chem. Eng.
Prog., 58, 40-44 (1962).
f Fluid jacket Edelmayer, A., Bokor, J., Szabo, Z., Szigeti, F.,
R Reactor Input reconstruction by means of system
W Wall inversion: a geometric approach to fault detection
A ambient and isolation in nonlinear systems, Int J Appl
Fe fluid inlet Math Comput Sci, 14, 189–199 (2004).
Fantuzzi, C., Simani, S., Beghelli, S., Rovatti, R.,
Greek Letters Identification of piecewise affine models in noisy
environment, Int J Control, 75, 18, 1472–1485
ρ Density kg.m-3 (2002).
Brazilian Journal of Chemical Engineering Vol. 25, No. 04, pp. 777 - 788, October - December, 2008
788 Y. Chetouani

Feil, B., Abonyi, J., Nemeth, S., Arva, P., Biochemical Engineering Journal, 14, 51-65 (2003).
Monitoring process transitions by Kalman Liberatore, S., Speyer, J. L., Hsu, A. C., Application
filtering and time-series segmentation, Computers of a fault detection filter to structural health
& Chemical Engineering, 29, 1423-1431 (2005). monitoring, Automatica, 42, 1199-1209 (2006)..
Forsythe, G. E., Malcolm, M., Moler, C., Computer Mehra, R. K., Peschon, J., An introduction approach
Methods for Mathematical Computations, to fault detection and diagnosis in dynamic
Prentice-Hall, New Jersey (1977). systems, Automatica, 5, 637-640 (1971).
Frank, P. M., Fault diagnosis in dynamic systems Nyberg, M., Stutte, T., Model based diagnosis of the
using analytical and knowledge-based air path of an automotive diesel engine, Control
redundancy. A survey and some new results, Engineering Practice, 12, 513-525 (2004).
Automatica, 3, 459-474 (1990). Patton, R. J., Frank, P. M., Clark, R. N., Issues of
García, F. P., Schmid, F., Conde, J., A reliability fault diagnosis for dynamic systems, Springer-
centered approach to remote condition Verlag, London (2000).
monitoring. A railway points case study, Pedregal, D. J., Carnero, M. C., State space models
Reliability Engineering & System Safety, 80, 33– for condition monitoring: a case study, Reliability
40 (2003). Engineering & System Safety, 91, 171-180
Henry, D., Zolghadri, A., Design of fault diagnosis (2006).
filters: A multi-objective approach, Journal of the Pedregal, D. J., García, F. P., Schmid, F., RCM2
Franklin Institute, 342, 421-446 (2005). predictive maintenance of railway systems based on
Himmelblau, D. M., Fault detection and diagnosis in unobserved components models, Reliability
chemical and petrochemical processes, Elsevier, Engineering & System Safety, 83, 103-110 (2004).
New York (1978). Shi, P., Liu, H., Stochastic finite element framework
Hovland, G. E., Von Hoff, T. P., Gallestey, E. A., for simultaneous estimation of cardiac kinematic
Antoine, M., Farruggio, D., Paice, A. D. B., functions and material parameters, Medical
Nonlinear estimation methods for parameter Image Analysis, 7, 445-464 (2003).
tracking in power plants, Control Engineering Simani, S. Fantuzzi, C., Dynamic system
Practice, 13, 1341-1355 (2005). identification and model-based fault diagnosis of
Isermann, R., Model-based fault-detection and an industrial gas turbine prototype, Mechatronics,
diagnosis – status and applications, Annual 16, 341-363 (2006).
Reviews in control, 29, 71–85 (2005). Tylee, J. L., On-line failure detection in nuclear
Isermann, R., Ballé, P., Trends in the application of power plant instrumentation, IEEE Trans
model-based fault detection and diagnosis of Automatic Control, 3, 406-415 (1983).
technical processes, Control Eng Pract, 5, 709– Villermaux, J., Génie de la réaction chimique-
719 (1997). conception et fonctionnement des réacteurs,
Izadi, I., Zhao, Q., Chen, T., An optimal scheme for Lavoisier, Paris (1993).
fast rate fault detection based on multirate Yang, S. K., An experiment of state estimation for
sampled data, Journal of Process Control, 15, predictive maintenance using Kalman filter on a
307-319 (2005). DC motor, Reliability Engineering & System
Jang, D-S., Choi, H-L., Active models for tracking Safety, 75, 103-111 (2002).
moving objects, Pattern Recognition, 33, 1135- Yang, S. K., Liu, T. S., State estimation for
1146 (2000). predictive maintenance using Kalman filter,
King, R., Gilles, R., Multiple filter methods for Reliability Engineering & System Safety, 66, 29-
detection of hazardous states in an industrial 39 (1999).
plant, AIChE, 11, 1697-1706 (1990). Zhan, Y. Makis, V., A robust diagnostic model for
Korbicz, J., Koscielny, J. M., Kowalczuk, Z., gearboxes subject to vibration monitoring,
Cholewa, W., Fault diagnosis: models, artificial Journal of Sound and Vibration, 290, 928-955
intelligence, applications (1st ed.), Springer- (2006).
Verlag (2004). Zogg, D., Shafai, E., Geering, H. P., Fault diagnosis
Li, J., Xu, N. S., Su, W. W., Online estimation of for heat pumps with parameter identification and
stirred-tank microalgal photobioreactor cultures clustering, Control Engineering Practice, 14,
based on dissolved oxygen measurement, 1435-1444 (2006).

Brazilian Journal of Chemical Engineering