Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 129

x7

Model-Based FDI Schemes For

Robot Manipulators Using Soft
Computing Techniques
Tolga YÜKSEL and Abdullah SEZGİN
Ondokuz Mays University
TURKEY

1. Introduction
While modern control methods were becoming widespread, in addition to demanded
repeatability and accuracy specifications, reliability and detection and isolation of probable
faults have become an obligation for automatic control systems. In the early 70’s, first
studies were appeared on this subject. While the first studies on fault detection and isolation
(FDI) were implemented for supervisory of chemical processes, following studies were
extended to systems like air and spacecrafts, automobiles, nuclear reactors, turbines and
HVACs with high reliability mandatories after especially aircraft accidents with high
mortality. In 1991, with extending and increasing studies, IFAC SAFEPROCESS comittee
was founded and in 1993, this comittee issued some definitions about fault types, fault
detection and isolation, fault diagnosis and fault tolerant control (FTC) (Isermann & Ballé,
1997).
Robots are accepted as an assistant subsystem or an individual part of a complex system in
most applications. In addition to applications like serial product lines in which they can
work harder, faster and with higher accuracy than humans, they are assigned to missions
like waste treatment in nuclear reactors, data and sample collection, maintenance in space
and underwater tasks which can be very risky for humans. As a consequence, a fault in one
product line may cause a pause in all connected lines even in flexible automation systems or
a developing and undetected fault may cause abortion of a whole space or underwater
mission with big money costs, it may even cause harm to humans. With the increase of these
events in real-life applications and with 90’s, studies on robot reliability and fault detection
and diagnosis in robotics have become common. In addition to these studies, NASA and US
Army issued some standards on robots and on the reliability and fault possibilities of
robotparts (Cavallaro & Walker, 1994).
This study is focused on model-based FDI schemes, how they can be applied to robot
manipulators, how soft computing techniques can be used in these schemes and three
different FDI schemes are proposed. Soft computing techniques which can overcome the
difficulties of schemes using analytical methods for nonlinear systems are used as
modelling, fault isolator and fault function approximator tools in the proposed schemes. In
the following section, a literature overview on FDI for nonlinear systems and robot

www.intechopen.com
130 Fault Detection

manipulators is given. In Section 3, defined faults and their physical causes are explained. In
Section 4, 5 and 6, the proposed schemes are introduced and explained. Soft computing tools
used in these schemes are introduced; furthermore, how they work and what their duties
are in these schemes are explained in details. In Section 7, simulation implementations and
results of these schemes for a two-link robot are given. In the last section, a comparison of
these schemes according to some FDI specifications and future studies on these schemes are
given.
Most studies in the literature are interested in sensor faults and locked and free-swinging
joint (actuator) faults. This study is interested in abrupt partial actuator faults defined in
Section 3 to contribute model-based FDI studies for robot manipulators. Furthermore, most
studies using soft computing techniques for FDI are interested in how the parameters in
their soft computing tools can be defined in terms of faults and how they can be updated
according to faults like adaptive learning. In this study, these tools are used directly and
without any modification to give appropriate outputs for appropriate inputs. From this
point of view, this study can be accepted as a bridge between model-based FDI and data-
based FDI methods. Furthermore, in this study, in addtion to soft computing tools, a hybrid
soft computing tool M-ANFIS (multiple-ANFIS) which combines and utilizes benefits of
neural network (NN) and fuzzy logic (FL) is used for modelling and function
approximation. The two-link robot manipulator used in the simulations can be seen simple
but accepted as a test platform for most studies. Besides, it has sufficient specifications for
implementation of newly proposed schemes.

2. Literature overview of model-based FDI for nonlinear systems and robot

manipulators
Studies and methods on fault detection and isolation can be divided into two main groups:
model-based methods and data-based methods (Chen & Patton, 1999). Model-based
methods are based on modelling the system and processing the difference signals between
the model and the real system named as residuals. Data-based methods are based on
processing the input and output signals of the system. Proposed schemes in this study are
based on model-based FDI methods and studies on model-based FDI are examined in
details. Information and surveys about data-based methods can be found in (Chen & Patton,
1999; Patton et al., 2000a; Venkatasubramanian et al., 2003). Model-based fault diagnosis
(detection and isolation) is defined as detection, isolation and characterization of faults in
components of a system from the comparison of the system’s available measurements, with
a priori information presented by the system’s mathematical model (Chen & Patton, 1999).
According to this definition, model-based FDI methods are formed of two steps. The first
step is generation of difference signals called residuals between real and predicted or
estimated output signals of the system. Discordance of these real and predicted or estimated
output signals which means nonzero residual signals indicates a potential fault in the
system. The second step is isolation of faults using these obtained difference signals
according to a decision set.
Model-based FDI methods can be classified according to the method used for residual
generation and the decision set is defined according to the specifications of each residual
generation method. The point of classification can not only be the method but also linearity
(linear-nonlinear-bilinear) type of the system that will be dealt with. The methods used for

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 131

linear systems can be classified into three main titles. The first and mostly used methods are
observer based methods. The main idea behind these methods is to estimate the output of
the system using the measurements with Luenberger observes and deterministic
adjustments or with Kalman filters and stochastic adjustments (Frank & Ding, 1997). Parity
vector (relation) methods use definition of parallel or temporal redundancy which is named
as defining a variable with two or more definitions and which can be obtained from
measurements or analytical relations (Chow & Willsky, 1984). Parameter estimation
methods are based on the principle that accepts sudden changes in parameters like friction,
mass, viscosity, resistance etc. using system identification methods as a sign of faults
(Isermann & Ballé, 1997; Moseler & Isermann, 2000). In addition to proposing new methods,
all these methods are investigated for the robustness against disturbances and uncertainties.
A detailed survey about these methods and studies can be found in (Chen & Patton, 1999;
Patton et al., 2000; Frank & Ding, 1997).
Generally, two main approaches are adopted for FDI for nonlinear systems (Chen & Patton,
1999). First approach linearizes nonlinear models around one or multiple
equilibrium(working) points and generates residuals insensitive to parameter changes in
small equilibrium point neighbourhoods using robust techniques. This approach may give
good and sufficient results only for systems with low level nonlinearities. But this approach
is not suitable for nonlinear systems having highly nonlinear terms and wide working
points. Second approach, as proposed for solving this problem, uses multiple nonlinear
models for each working points. But this approach bring multiple FDI systems for each
working points and this will be not practical for real-time implementations.
As mentioned below, to solve these problems, FDI methods which can deal with nonlinear
systems directly must be proposed and developed. Therefore, most of the methods
proposed for linear systems are adapted for nonlinear systems. One of the approaches is to
use analytical or deterministic nonlinear observers (Chen & Patton, 1999; Patton et al. 2000;
Frank & Ding, 1997; Adjallah et al.,1994; Garcia & Frank, 1997). Model of a nonlinear
systems is accepted as below:

x (t )  g( x(t ), u(t ), f (t ), d(t ))

y(t )  h( x(t ), u(t ), f (t ), d(t ))
(1)

where x(t) is the state vector, y(t) is the output vector, u(t) is the input vector, f(t) is the fault
vector, d(t) is the disturbance vector and g( . , . , . , . ) and h( . , . , . , .) are nonlinear functions.
FDI problem is generating residuals using the observer form defined in (2):

(t )  gr ( (t ), u(t ), y(t ))

r (t )  hr ( (t ), u(t ), y(t ))
(2)

and the residuals should prove (3):

0 f (t )  0
r (t ) 
 0 f (t )  0
(3)

Here, the aim is to design ξ appropriate for gr( . , . , . ) and hr( . , . , . ). ξ(t) defines state

www.intechopen.com
132 Fault Detection

estimator (Garcia & Frank, 1997). In the literature, there are some nonlinear observer desing
approaches defined for certain type nonlinear systems (Frank & Ding, 1997; Adjallah et al.,
1994; Garcia & Frank, 1997; Seliger & Frank, 1991; Yang & Saif, 1995; Kinnaert, 1999).
All approaches mentioned below use analytical methods but it is hard to obtain analytical
models which nonlinear observers are based on. To overcome this problem, “universal
approximator” soft computing modelling tools which can model nonlinear systems are
preferred. Soft computing techniques involves NN, FL and genetic algorithms (GA). NN can
be used to model multiple-input-multiple-output (MIMO) nonlinear systems by using
nonlinear mapping capabilities in its hidden black-box structure (Haykin, 1999). Residual
generation is implemented by comparing the real system outputs and estimated outputs by
NN. Furthermore NNs are superior to analytical techniques on classification and a second
NN can be used to isolate faults by evaluating (classifying) the residuals (Patton et al.,
2000b; Marcu et al., 1998). Model-based FDI scheme with NNs is shown in Fig. 1:

Fig. 1. Model based fault detection and isolation with two neural networks

In addition to these NN approaches, there are some approaches named online approaches
using adaptive learning and defining NN parameters in terms of input signals anf fault
functions (Zhang et al., 2002; Polycarpou & Helmicki, 1995; Polycarpou & Trunov, 2000)
NNs with self learning capabilities can be thought convenient for FDI but, as a black-box
structure, keeping user experiences and interferences away from modelling is not a desired
feature. To overcome this disadvantage, FL is used as modelling, observer, adaptive residual
threshold selector tools for FDI for nonlinear systems (Dexter, 1995; Patton et al., 1998; Ballé,
1998). The main idea behind fuzzy observers is to define a nonlinear system using a set of
locally linearized observers with Takagi-Sugeno fuzzy models. Local region definitions are
generated using working points. Furthermore, FL is used for residual evaluation. The most
important problem in residual evaluation is to set residual threshold. An adaptive threshold
can prevent the FDI system from false and missed alarms and FL can be used as the
adaptive threshold selector tool (Schneider & Frank, 1996).
FDI studies on robots continue on most of the robot types (mobile, flexible, kinematically
redundant, parallel, mobile manipulator, humanoid, bipedal and multi-legged etc.). This
study deals with the most common used serial, open chained and rigid robot manipulator
and only a detailed review on this type of robots is given here. Studies on other type robots

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 133

can be found in (Goel et al., 2000; Tinós & Terra, 2002). Most studies on FDI for robot
manipulators are based on nonlinear observer approaches. Caccavale and Walker tried to
adapt robot dynamics to a certain kind of nonlinear systems and used nonlinear observers
which are convenient for these systems (Caccavale & Walker, 1997). Similarly, Schneider
and Frank used a robust observer designed for nonlinear systems for robot dynamics and
fuzzy logic for residual evaluation/fault isolation (Schneider & Frank, 1996). Leuschen et al.
transferred analytical redundancy from linear to nonlinear systems, they defined nonlinear
analytical redundancy term and they implemented these redundancies and FDI studies on a
hydraulic robot manipulator platform and on a two-link IMI robot manipulator (Leuschen
et al., 2005). De Luca and Mattone resembled robot dynamics to generalized notation of
adaptive controllers and proposed an adaptive FDI scheme (De Luca & Mattone, 2004).
Dixon et al. proposed a filter, they passed torque signals applied to a robot manipulator
through this filter and they proposed an FDI technique robust to parametric uncertainties
using the difference signals between these filtered signals and their predictions (Dixon et al.
2000). Abdul and Liu proposed an analytical method for the prediction of position and
velocity signals of modular type robots and designed a fault tolerant controller using these
predictions (Abdul & Liu, 2008). Brambilla et al. generated residuals using inverse robot
model, proposed a sliding mode observer and isolated sensor faults using this observer
(Brambilla, 2008). Chen and Saif resembled robot dynamics to systems with unknown
inputs in state-space and implemented fault detection using output observers (Chen & Saif,
2008).
NNs as an approved tool for FDI for nonlinear systems are also used for robot manipulators.
Naughton et al. used nonlinear observer proposed by Adjallah et al. for residual generation
and NN for residual evaluation (Adjallah et al., 1994; Naughton et al., 1996). Vemuri and
Polycarpou considered fault as a component of robot model function and used adaptive
learning strategy of NNs to approximate fault function (Vemuri & Polycarpou, 1997). Terra
and Tinós used some different types of NN structures for both residual generation and
evaluation (Terra & Tinós, 2001). Lee et al. tried to use parameter identification methods for
fault detection and ART type NNs for fault isolation on component and sensor type faults
(Lee et al., 2003). Datta et al. tried to classify coefficients obtained from discrete wavelet
transform (DWT) using a NN (Datta et al., 2007).

3. Faults defined for robot manipulators

Faults can be classified according to the part of the system, according to modelling or
according to time characteristics (Chen & Patton, 1999). In this section, how faults defined
for robot manipulators are classified according to the part of the system (Fantuzzi et al.,
2003).
Generalized dynamics of robot manipulators are defined in (4) :
M ( q )q  V ( q , q )  G(q )  F(q )   (4)

In (4), n as the number of links, q(t ), q (t ), q(t )  n are in order angular position, velocity and
accelerations of each link, M ( q )   nn is the positive defined inertia matrix,V (q , q )  n is the
Coriolis and centripedal vector, G ( q )   n is the gravity vector, F ( q )   n is the friction
vector and    n is the vector of applied torques to joints. If the nonlinear terms in (4)
except terms having angular acceleration are expressed as N ( q , q )  V ( q , q )  G( q )  F ( q ),

www.intechopen.com
134 Fault Detection

(4) becomes

M (q )q  N (q , q )   (5)

Faults are defined in (5) as follows:

a) Actuator faults: Motors and power transmission tools are charged as actuators for
manipulators and faults occured in these actuators affects the ability of movement. These
type of faults change dynamics as follows:

M(q )q  N (q , q )  u(t  T ) f (t )   (6)

• Locked joint faults: These faults occur when the magnetic brake of the motor connected to
a joint is locked and doesn’t allow any movements (τ free - qi fixed).
• Free-swinging joint faults: These faults occur when applied torque of the motor connected
to a joint is zero because of a disconnected cable (τ zero - qi free and under impact of other
joints or gravity).
• Partial actuator faults: These faults occur when applied torque of the motor connected to a
joint decreases (i.e. %20 decrease etc.) because of a fault at power electronics components.
b) Component faults: In robot manipulators, broken link, gear corrision or fixed or slipped
chain can be component faults. Dynamics of the manipulator change as follows:

M(q )q  ( N (q , q )  f N )   (7)

c) Sensor faults: Optical encoders for joint positions, tachogenerators for joint velocities and
tactile sensors for contact forces are the sensors used in robot manipulators. The faults
defined for all sensors can be seen in these sensors (bias etc.). Dynamics of the manipulator
change as follows:

y(t )  [ q1 (t )  qn (t ) q 1 (t )  q n (t )]T  f q ,q (8)

In (6), (7) and (8), u(t-T) is the delayed unit step function, T is the fault occurence instant, fN
is the component fault, f q , q is the sensor fault.

4. FDI scheme with M-ANFIS and NN

The block diagram of the first scheme is shown in Fig. 2. The diagram includes not only the
FDI block but also the control block to expose the whole system. To control robot,
Computed Torque-PID (CT-PID) method is used and a brief introduction about this method
is given in Section 6. CT-PID method needs exact robot model but it can cope with bounded
uncertainties. M-ANFIS as the first block of FDI scheme is formed of multiple independent
ANFIS structure and implements robot modelling. It takes instant torque values of each
joint and unit time step delayed angular position and velocity values of each link. The
model gives estimated values of instant angular position and velocity values of each link.
Difference signals between real robot and robot model generate residuals. Fault detection
alarm is accepted as overshooting of predefined thresholds by residuals obtained from

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 135

healthy robot operations or simulations. Fault isolation process is implemented by

classification of the generated residuals applied to a multilayer NN. In subsections, a review
on M-ANFIS and NN with resilient propagation will be given.

Fig. 2. The block diagram of the first FDI scheme

4.1 Residual generation with M-ANFIS

Difficulties in modelling nonlinear systems analytically, robustness obligations to
uncertainty and disturbance directed researchers to soft computing techniques and
structures with self, automatic learning and nonlinear mapping capabilities. Soft computing
techniques involve NNs, FL, GA and hybrid structures of these tools. Despite NNs have lots
of types and lots of learning algorithms, they behave like a black box due to their self-
learning nature. Alike FL leaves all parametric adjustments to users, users’ experiences
become a parameter in modelling and performance of adjustments are dependent on the
users. Jang et al. considered advantages of both structures and they decided to combine
these advantages and proposed an adaptive network called ANFIS (Adaptive Neuro Fuzzy
Inference System) which is functionally equivalent to fuzzy inference system (Jang et al.,
1997). ANFIS has a 5-layered structure and a sample with 2 inputs is shown in Fig. 3.
Functions of the layers are given below.

Fig. 3. General ANFIS structure

Layer 1: This layer contains membership functions of inputs as defined in (9) and all inputs
are applied to these functions. Type and shape of the membership functions are defined by
the user and generally, these functions are bell-shaped functions defined in (10):

www.intechopen.com
136 Fault Detection

L1xi   Ai ( x ) i  1,2
L1yi  Bi ( y ) i  1,2
(9)

 Ai ( x )  ,  Bi ( y )  i  1,2
1 1
(10)
x  c Ai y  c Bi
2 b Ai 2 bBi
1 1
a Ai aBi

where ai is center, ci is width and bi is crossover gradient, (ai, bi, ci) are parameters of defined
function and named as premise parameters.
Layer 2: Each function value is multiplied by other values coming from other inputs due to
defined rule base and the result values are named as firing strength of each rule:

L 2 i  w i   Ai ( x ). Bi ( y ) i  1,2 (11)

Layer 3: Firing strengths are normalized:

L 3i  wi  i  1,2

w
wi
2
(12)
j
j 1

Layer 4: Normalized firing strengths are multiplied by a first order function of inputs:

L 4i  wi . f i  wi ( pi x  qi y  ri ) i  1,2 (13)

where (pi, qi, ri) are parameters of a first order function and these parameters are named as
consequent parameters.
Layer 5: Values coming from all Layer 4 outputs are summed and output value is obtained.

w f
2
L5  i i (14)
i 1

Points that use user experience in ANFIS like in FL are the choices of membership function
types at Layer 1 and multiplication operation due to rule base arrangement at Layer 2. These
dependencies are exposed in the demonstration of functional equivalence of ANFIS and
fuzzy inference system under some circumstances in (Jang et al., 1997).
It is purposed in all network structures to adapt or update network parameters in order to
give appropriate outputs against appropriate inputs. From this point of view, ANFIS
updates its own parameters using learning algorithms like NNs. As a learning algorithm,
backpropagation or hybrid learning expressed in forward and backward passes can be
preferred. Table 1 explains hybrid learning. In forward pass, while premise parameters are
fixed, inputs go forward until Layer 4 and consequent parameters are determined with least

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 137

squares. In backward pass, consequent parameters are fixed, error is backpropagated until
Layer 1 and premise parameters are determined with gradient descent.

Parameters Forward pass Backward pass

Premise Fixed Gradient descent

Consequent LSE Fixed

Table 1. Parameter updates for hybrid learning in two passes

Fig. 3 illustrates the main disadvantage of ANFIS, being multi-input-single-output (MISO).

To model systems with multiple outputs (MIMO), multiple-ANFIS (M-ANFIS) which has
independent parameters and outputs is used. A structure that considers correlations
between outputs and uses some mutual parameters in order to decrease computational load
coming from increasing parameter number exists and it is named as coactive-ANFIS (C-
ANFIS) (Jang et al., 1997). In this study, M-ANFIS is preferred and M-ANFIS with 2 inputs
and 2 outputs is illustrated in Fig. 4.

Fig. 4. M-ANFIS with 2 inputs-2 outputs

Robot model takes instant torque and unit time step delayed position and velocity signals
belonging to each joint as inputs and generates instant predictions of position and velocity
signals of each joint using an ANFIS for each signal. Residuals are generated from
differences between real robot and robot model signals. Residual generation using M-ANFIS
is shown in Fig. 5.

Fig. 5. Residual generation with M-ANFIS

www.intechopen.com
138 Fault Detection

If one of the residual signals overshoots the thresholds determined by simulations and
implementations running under healthy conditions, it informs a fault alarm. In this study, in
addition to generated residuals, signals called analytical redundancies are used for fault
detection and isolation. These signals will be outlined in the following section.

4.2 Residual evaluation using NN with resilient propagation and analytical redundant
signals
Residual generation is followed by residual evaluation to isolate faults. This operation is
based on the fact that different types of faults show different residual characteristics and it is
considered as a classification process. All techniques used for classification and pattern
recognition can be used for this process. Multilayer feedforward NNs are very convenient in
soft computing based classification tools (Haykin, 1999). (Haykin, 1999) gives all detailed
information about NNs. Here, just a review on the used learning algorithm will be given.
NNs aim to update their defined parameters with learning algorithms to give appropriate
outputs for appropriate inputs. The most common learning algorithm for this purpose is
gradient descent algorithm and it is defined in (15):

wij( n  1) wij ( n)  wij (n)

E(n)
wij (n)   .
(15)
wij

where E(n) as the n. step error function, wij as the weight from neuron i to neuron j and η as
the learning rate parameter. The updates are dependent on the learning rate parameter η
and it is known that if it is chosen too small, too many operation steps will be needed and if
it is chosen too large, minimum value will not be reached and error value will oscillate
around it. To avoid this problem and to accelerate converge, a momentum term with μ
momentum parameter is added to (15):

E(n)
w ij (n)   .   .w ij (n  1)
w ij
(16)

Despite the momentum term in (16), observations showed that this regulation is still
dependent on the selected momentum parameter. And again to avoid these parameter
dependencies, adaptive learning and momentum parameters are suggested.
These regulations neglect that weight updates are dependent not only the learning rate but
also partial derivatives of E(n) with respect to wij. Resilient Propagation (RP) learning
algorithm removes this blurred adaptation from updates and performs updates directly
according to the following steps (Riedmiller & Braun, 1993). Firstly, RP assigns Δij update
values to all weights. These values are updated as defined in (17) and (18):
(17) expresses that if the partial derivative of error with respect to weight changes its sign,
the update value is too big and local minima is missed, it should be decreased by η− factor
taking values between 0 and 1 and if it remains with the same sign, it should be increased
by η+ factor taking values greater than 1. After the update value is calculated, the weight
update is performed in (18):

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 139

  ( n  1) E( n  1) E(n)
 . ij if
wij
.
wij
0

  ( n  1) E( n  1) E(n)
 . ij 0
 ij (n)   wij wij
if .
(17)
 ( n  1)
  ij else

 0    1  
 

 ( n ) E(n)
  ij if w  0

 ( n ) E(n)
ij

wij( n )    ij if 0
wij
(18)

0

else


On the contrary of other learning algorithms, RP is more transparent, has a more powerful
update process and is more efficient with respect to time, memory storage consumption and
it is chosen as the learning algorithm in this study.
Simulation results in this study and in some other studies showed that it is hard to train a
NN with sufficient fault isolation rates just using existing residual signals (Chen & Patton,
1999; Leuschen et al., 2005). To assist isolation process, some variant signals must be
generated. This can be done by using the definition of analytical redundancy as defining
one variable in two or more ways like derivation of position and integral of acceleration for
velocity. It is clear that the derivative value of position residuals must be equivalent to
velocity residuals mathematically and these derivatives can be applied to NN for isolation
process and can be used for fault detection. It must be noted that these redundant signals
are just used to help fault isolation (classification) process and these signals are not directly
sensitive to a fault defined in (Leuschen et al., 2005). Definitions of analytical redundant
signals and fault detection process in this study are given in (19) and (20), respectively. NN
for fault isolation is shown in Fig. 6.

( rq (t ))  ( q(t )  qˆ (t ))

 q (t )  qˆ (t )
(19)

rq (t )  threshold _ uprq
rq (t )  threshold _ downrq
rq (t )  threshold _ uprq
rq (t )  threshold _ downrq  fault alarm (20)
( rq (t ))  threshold _ up( rq )
( rq (t ))  threshold _ down( rq )

www.intechopen.com
140 Fault Detection

Fig. 6. Fault isolation with NN using residuals and analytical redundancy

5. FDI scheme using generalized observers with M-ANFIS

The design procedure of fault isolation using observes is based on defining relationships
between faults and generated residuals. If the residual set can isolate all faults, it can be said
that the residual set has the required isolation property. Two methods can be applied to
residual sets that involve all residuals to verify this fault isolability property (Chen & Patton,
1999). For dedicated observer schemes (DOS), as the first method, each of the residuals must
be sensitive to one fault and insensitive to others. Although this method sounds good, in
practice, it is hard to design and to obtain robustness against modelling errors for dedicated
observers using analytical approaches. For generalized observer schemes (GOS), as the
second method, each of the residuals is sensitive to all but one fault and it is easier to design
generalized observers using generated residuals. GOS approach can easily cope with
uncertainties by modelling faulty systems one by one. Nevertheless, computational load
coming from each system models arises real-time implementation problem. The block
diagram of the second proposed model-based FDI scheme is shown in Fig. 7.

Fig. 7. The block diagram of the second FDI scheme

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 141

Again CT-PID is preferred as the robot control method. For the robot manipulator, n+ 1
systems, as one healthy and n faulty, are defined. Each system is modelled using M-ANFIS
and generates residual sets (position and velocity residuals of each system) using the same
definitions in Section 3.1. Using GOS approach, each of the residual sets overshoots
thresholds of all other systems but itself determined by simulations or implementations.
GOS combines fault detection and isolation process in one step and that makes the method
more effective and attractive. But using models for each defined fault exposes heavy
computational load. Table 2 illustrates residuals and residual evaluation process for the
proposed scheme.

Faults r1 r2 r3 … rn+1
Healthy 0 1 1 1 1
f1 1 0 1 1 1
f2 1 1 0 1 1
… 1 1 1 0 1
fn 1 1 1 1 0
Table 2. Residual evaluation for generalized observer scheme

6. Fault function approximator FDI scheme with M-ANFIS

Despite two proposed schemes are valid and effective for fault detection and isolation, due
to their nature, they can be used only for predefined fault types and that makes them hard
against partial actuator faults for nonlinear systems and robot manipulators. Furthermore,
these schemes can not fully succeed in the design of fault tolerant controllers that use
information coming from FDI.
These disadvantages makes fault function approximation more important and the third
proposed scheme is focused on fault function approximation. The block diagram of the third
scheme is shown in Fig. 8. Modelling part of the scheme is implemented using the same
definitions in Section 3.1. In the design of this scheme, to approximate faults, firstly, a M-
ANFIS that takes residuals and redundant signals as inputs is trained but the results are
vain. As the second attempt, a M-ANFIS that uses instant torque, position and velocity
signals as inputs is trained. Each ANFIS of M-ANFIS gives a fault function approximation of
each actuator belonging to each joint. Activation of M-ANFIS and fault detection process are

Fig. 8. The block diagram of the third FDI scheme

www.intechopen.com
142 Fault Detection

implemented by residuals and analytical redundant signals. M-ANFIS is designed to give

one fault function, the other outputs will give zero value. That makes the fault isolation
process unnecessary but simulation results showed that especially for small faults more than
one output may give approximations. In this study to avoid this problem, bigger output is
accepted as the real fault function approximation.

7. Simulation results
In this section, the proposed FDI schemes are simulated using MATLAB Fuzzy Logic Toolbox
and Neural Network Toolbox. Robot manipulator used in the simulations is a two-link planar
manipulator under gravity and masses of the links are defined at the end of the links (Lewis
et al., 1993). The manipulator is shown in Fig. 9.

Fig. 9. Two-link manipulator under gravity

The dynamics of the manipulator and the generalized form of dynamics are given in (21)
and (22), respectively:

M11  (m1  m2 )a12  m2 a2 2  2 m2 a1 a2 cos(q2 )

M12  M 21  m2 a2 2  m2 a1 a2 cos(q2 ) , M 22  m2 a2 2
V1  m2 a1 a2 ( 2 q 1q 2  q 2 2 ) sin(q2 )
V2  m2 a1 a2 q 12 sin(q2 ) (21)
G1  (m1  m2 ) ga1 cos(q1 )  m2 ga2 cos(q1  q2 )
G2  m2 ga2 cos(q1  q2 )
 M11 M12  V1  G1 
M(q )    , V (q , q )    , G(q )   
 M 21 M 22  V2  G2 

  M (q )q  V (q , q )  G(q )  M (q )q  N (q , q ) (22)

where    2 is the applied torques to joints, M ( q )   2 2 is the inertia matrix, V ( q , q )   2

is the Coriolis/centripedal vector and G ( q )   2 is the gravity vector. Friction and
disturbance terms are neglected. The link masses are m1 = m2 = 1 kg., the link lengths are
a1 = a2 = 1 m. and the sampling frequency is 100 Hz.

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 143

Computed Torque control is a special application of feedback linearization of nonlinear

systems to robotics. It has commonly PD or PID types and CT-PID is used in this study
(Lewis et al., 1993). CT-PID is not very effective against uncertainties but in this study, the
controller is not important because suggested FDI schemes use soft computing techniques
and they use only the datas coming from the system. If another controller is used, only datas
will change and soft computing tools will be trained again using these new datas. Equations
of CT-PID are given in (23):

  e , e  q d  q
  M (q )(qd  K d e  K p e  K i )  N ( q , q )
(23)

where e is the error, ε is the derivative of error. The gain matrices of the controller are Kd =
100 I2×2, Kp = 20 I2×2, Ki = 500 I2×2.
In this study, 4 different partial actuator faults are defined, examined and they are given in
Table 3 with loss percentages. Expression of partial actuator faults defined in (6) with
respect to time in (22) is given in (24).

M (q(t ))q(t )  N (q(t ), q (t ))   (t )   .u(t  T ). (t ) (24)

where   2 is the partial actuator loss, u (t )   2 is the unit step and T is the fault
occurence instant.

Fault Name

Actuator 1 %50 loss f1

Actuator 1 %30 loss f2
Actuator 2 %50 loss f3
Actuator 2 %30 loss f4

Table 3. Simulated faults

7.1 Case 1: FDI with the first scheme

As the first case, FDI scheme in Fig. 2 is implemented. Firstly, M-ANFIS is constructed as
the model. M-ANFIS is formed of 4 independent ANFIS having instant torque and unit time
step delayed position and velocity signals as common 6 inputs coming from 2 joints and
each giving assigned instant position or velocity signal of each joint. Manipulator is
simulated for 146 different joint trajectories defined in sin-, cos-shape with amplitudes
varying between ±1 and 76 sampled datas from these simulations are used to train M-
ANFIS. Each ANFIS has two bell-shaped membership functions for each input and hybrid
learning is selected as the learning algorithm. To show the robustness of the model and
residuals generated, time varying %5 dynamics uncertainty is added to (22) as accepted in
(25) and the schemes are tested against different uncertainties added to robot dynamics.

www.intechopen.com
144 Fault Detection

M(q )  1.05 M (q ) , N (q , q )  1.05 N (q , q ) (25)

Difference signals between this robot model and real robot generate residuals. Although
residuals must be zero in healthy conditions, it may have very small values due to
uncertainties. 12 trajectories of 146 trajectories are selected, up and down boundary
thresholds of residuals generated from healthy 12 different trajectories are determined and
any overshooting is defined as the fault alarm and fault detection part of the scheme.
A four-layered NN with 10-20-20-4 neurons using RP learning algorithm is constructed for
residual evaluation/fault isolation. Four faults defined in Table 4 for the same 12 different
trajectories are simulated, 4 residuals and 2 analytical redundant (position residual
derivatives of each link) signals with 101 samples (6×4848) as 6 inputs and 4 outputs giving
0.9 value, each representing one fault in Table 4, are used to train NN. 7000 epochs and 0
error target are selected as training parameters. Training process is accomplished under 5
minutes and error value is under 10-3.
Simulation tests showed that NN is confused by residuals during transition from healthy to
faulty condition and this increases false alarm rate. To prevent this, 3.85 seconds delay is
added before NN starts to evaluate residuals after fault detection alarm.
If one of the NN outputs overshoots 0.5, it is defined as fault isolation signal. Simulation
studies showed that some faults belonging to some trajectories may cause two outputs
(especially fault outputs of the same actuator) to overshoot 0.5 for short time durations, and
these results increase false alarm rate. To regulate this, continuity is accepted as criteria and
network output signals with time durations shorter than 0.35 s. are neglected.
With these specifications, simulation studies showed %89.58 fault isolation rate for defined
faults and trajectories.
As an illustration of the proposed scheme, a trajectory is defined in (26), a %50 fault at
actuator 1 (f1) at t = 24 s. is simulated. Followed trajectories by joints, errors and applied
torques to joints are given in Fig. 10, residuals and analytical redundant signals are given in
Fig. 11 and NN outputs are given in Fig. 12 for 4-50 s. time interval:

qd 1 (t )  0.8 * cos(t / 2 )
q d 2 (t )  0.8 * sin(t / 2 )
(26)

It can be seen in Fig. 12 that the outputs of f1 and f2 overshoots 0.5 but the time continuity
rule neglects this time interval. And fault isolation instant t = 28.21 s. can be seen in Fig. 12.

7.2 Case 2: FDI with the second scheme

As the second case, scheme in Fig. 7 is implemented. Four actuator faults, as in the first case
defined in Table 1, are considered and from this consideration, as can be seen in Fig. 7, 1 M-
ANFIS for healthy and 4 M-ANFIS for faulty robot models are defined and each is formed
of 4 independent ANFIS having instant torque and unit time step delayed position and
velocity signals as common 6 inputs and each giving assigned instant position or velocity
signal. As a result, 5 M-ANFIS with 20 independent ANFIS is constructed. To train, same
146 different joint trajectories defined in sin- and cos-shape with amplitudes varying
between ±1 are simulated for healthy and faulty robot models and 76 sampled datas from
these simulations are used.

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 145

a)

b)

c)

Fig. 10. Time histories of joint 1,2: a) Followed trajectories b) Errors c) Applied torques
a) b)

c) d)

e) f)

Fig. 11. Residuals and analytical redundant signals with thresholds

a) r1  q 1  q̂ 1 b) r2  q 1  q̂ 1 c) r3  q 2  q̂ 2 d) r4  q 2  q̂ 2 e) r1  q 1  q̂ 1 f) r3  q 2  q̂ 2

www.intechopen.com
146 Fault Detection

Fig. 12. Neural network outputs

The operation of GOS shows that it can detect and isolate faults at the same time instant.
Although Table 2 implies that the residual of the defined fault or healthy condition must be
zero, it may have very small values due to modelling errors. To define thresholds for each
residual of defined conditions (1 healthy + 4 faulty), 80 trajectories are simulated, up and
down thresholds are determined, overshooting of these thresholds for each residual is
defined as another condition except belonging to itself. Analytical redundant signals are not
defined for this scheme because the residual signals are sufficient and capable of isolating
faults using GOS scheme.
Simulation studies for determining fault isolation rate showed that for some faults
belonging to some trajectories, two residual sets coming from two models may fall between
threshold intervals causing decrease of isolation rate. To increase this rate, residuals are
evaluated in the sequence of healthy-f1-f2-f3-f4. Furthermore, again continuity is accepted as
criteria and residuals falling between threshold intervals with time durations shorter than
0.8 s. are neglected.
With these specifications, simulation results showed %87.81 fault isolation rate for defined
faults and trajectories. As an example of this scheme, the same trajectory in (26) and the
same fault f1 is accepted. Residuals of f1 are given in Fig. 13 and isolation signals are given in
Fig. 14. Fault isolation instant t = 24.93 s. can be seen in Fig. 14.

7.3 Case 3: FDI with the third scheme

As the third case, the scheme in Fig. 8 is implemented. Robot modelling, residual generation
and fault detection processes are implemented in the same way with the first scheme. The
scheme is simulated for the 80 trajectories used in the second scheme and residual
thresholds are determined. To approximate the fault function, M-ANFIS is formed of two
independent ANFIS (for each actuator) taking instant torque, position and velocity signals
as inputs. Activation of M-ANFIS is realized by the residuals and analytical redundant

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 147

a) b)

c) d)

Fig. 13. Residuals of f1 : a) r1  q1  qˆ1 f1 b) r2  q1  qˆ1 f1 c) r3  q 2  qˆ 2 f1 d) r4  q 2  qˆ 2 f1

Fig. 14. Fault isolation signals of the second scheme

www.intechopen.com
148 Fault Detection

signals defined for the first scheme. For the training of M-ANFIS, to realize fault function
approximation more correctly, 2 extra faults f5-Actuator 1 %70 torque loss and f6-Actuator 2
%70 torque loss are defined. Simulations are implemented for defined 6 faults and 101
sampled datas from these simulations are used for training.
Function approximator M-ANFIS is formed of two ANFIS and they are trained to give fault
approximation of the faulty actuator. It is hard for the other ANFIS to give zero output
because of uncertainties and that causes false alarms. This problem is solved by the rule
“The bigger signal is the right signal”. This rule is based on that ANFIS output torque of the
faulty actuator is mostly bigger than other ANFIS output torques. This rule constitutes false
alarms in the cases of low percentage actuator faults resulting close ANFIS outputs.
The second problem is the torque signals containing high frequency components coming
from the change of the system region from stable to unstable in the cases of some very high
percentage actuator faults. Sudden changes in short time intervals as the high frequency
signals lead M-ANFIS to false alarms. These high frequency signals are removed by using a
filter like sliding mode control with saturation but the problem still continues and these
high frequency components are accepted as a part of the study.
In the simulation studies, like the first scheme, transition from healthy to faulty condition
increased false alarms and to decrease these alarms, function approximator M-ANFIS is
activated after 3.2 s. delay from fault detection instant.
Again for this scheme, 0.8 s. continuity test is added to the scheme to increase fault isolation
rates.
With these specifications, simulation results showed %87.5 fault isolation rate for defined
faults and trajectories. Fault function and aproximation signals are given in Fig. 15 and fault
isolation signals are given in Fig. 16 for the same example defined for the other schemes.
Fault isolation instant t = 28 s. can be seen in Fig. 16. In addition to approximating fault
function, this approximation can give information if the fault function causes actuator
saturation problem.

Fig. 15. Fault function and approximation signals

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 149

Fig. 16. Fault isolation signals of the third scheme

8. Comparison of the schemes and conclusions

In this study, three model-based fault detection and isolation schemes for robot
manipulators using soft computing techniques are proposed. All schemes use multi-input-
multi-output type of ANFIS, M-ANFIS structure for modelling. The first scheme uses
overshooting of the residuals for fault detection and a NN for fault isolation and the second
scheme uses generalized observer approach for fault detection and isolation. The third
scheme differs from these schemes by approximating fault function and it uses M-ANFIS for
approximation. A comparison of the schemes according to some important FDI
specifications is given in Table 4.
All schemes have low false alarm rates, common and positive responses against robustness,
and adding new trajectories and training. It is hard to define robustness of these schemes
analytically but all schemes are tested against different uncertainties added to robot
dynamics and all give similar isolation results. It is observed that new added position
trajectories with lower values than ±1 peak values show same results. When new trajectories
are added, residual thresholds should be calibrated, isolation structure should be trained
and isolation delay and continuity time intervals should be revised if the same isolation
results are desired.
One of the most demanded specification for FDI schemes is low or absent fault detection
and isolation delays. None of the schemes need fault detection delay. The first and the third
schemes have fault isolation delays arising from transition from healthy to faulty condition
and these delays should be paid attention especially in real-time implementations. The
second scheme realizes detection and isolation in the same method and doesn t need any
delays.

www.intechopen.com
150 Fault Detection

Resembling of faulty condition residuals to healthy and other type residuals in short time
intervals causes low isolation rates. To prevent this, continuity tests are added to all of the
schemes.
Fault Adding New
False Alarm Fault Isolation
Scheme Robustness Detection Trajectories&
Rate Delay
Delay Training
1. Scheme low reliable no yes easy

2. Scheme low reliable no no easy

3. Scheme low reliable no yes easy

Continuity Computational Fault Function
Usable For Fault Tolerance
Test Load Approximation
just defined
1. Scheme yes low faults just defined faults
just defined
2. Scheme yes very high just defined faults
faults
3. Scheme yes low yes yes
Table 4. Comparison of the schemes

The first and the third schemes can be considered to have low level, the second scheme to
have very high level computational load (simulation durations as the proof). The main
sources of the computational load in these schemes are soft computing structures.
Especially, ANFIS and its big brother M-ANFIS bring high computational loads. The second
scheme includes M-ANFIS for all models and that makes it hard for real-time
implementations.
The main disadvantage of the first and second scheme is that both are functional for just
defined faults. This is a limiting characteristic for fault function approximation and for fault
tolerant controller (FTC) schemes as an extension of FDI schemes. Fault function
approximation is important to give information about how fault function changes and to
give the peak value of control signal to observe the saturation occurence. The third scheme
can approximate all faults and that makes it the first choice for real-time implementation.
These three schemes, especially the third one, are suitable for real-time implementation. In
the future studies, schemes involving fault tolerant controllers will be proposed. These FTC
schemes will use gain scheduling nonlinear PID approach and all proposed schemes will be
implemented on a real manipulator.

9. References
Abdul, S. & Liu, G. (2008). Decentralised Fault Tolerance and Fault Detection of Modular
and Reconfigurable Robots with Joint Torque Sensing, IEEE International Conference
on Robotics and Automation, pp. 3520-3526
Adjallah, K.; Maquin, D. & Ragot, J. (1994). Non-linear observer-based fault detection, Third
IEEE Conf. on Control Applications, pp. 1115-1120

www.intechopen.com
Model-Based FDI Schemes For Robot Manipulators Using Soft Computing Techniques 151

Ballé, P. (1998). Fuzzy Model-Based Symptom Generation and Fault Diagnosis for Nonlinear
Processes, IEEE International Conference On Fuzzy Systems, Vol. 2, pp. 945-950
Brambilla, D.; Capisani, L. M.; Ferrara, A. & Pisu P. (2008). Fault Detection for Robot
Manipulators via Second-Order Sliding Modes , IEEE Transactions on Industrial
Electronics, Vol. 55, No. 11, pp. 3954-3963
Caccavale, F. & Walker, I. D. (1997). Observer-based fault detection for robot manipulators,
IEEE International Conference on Robotics and Automation, pp. 2881-2887
Cavallaro, J. R. & Walker, I. D. (1994). A survey of NASA and military standards on fault
tolerance and reliability applied to robotics, AIAA
Chen, J. & Patton, R. J. (1999). Robust Model-Based Fault Diagnosis For Dynamic Systems,
Kluwer Academic Publishers
Chen, W. & Saif, M. (2008). Output Estimator Based Fault Detection for a Class of Nonlinear
Systems with Unknown Inputs, American Control Conference, pp. 3307-3312
Chow, E. Y. & Willsky, A. S. (1984). Analytical redundancy and the design of robust failure
detection system, IEEE Transactions on Automatic Control, Vol. 29 No. 7, pp. 603-614
Datta, A.; Mavroidis, C.; Krishnasamy, J. & Hosek, M. (2007). Neural network based fault
diagnostics of industrial robots using wavelet multi-resolution analysis, American
Control Conference, pp: 1858-1863
De Luca, A. & Mattone, R. (2004). An adapt-and-detect actuator FDI for robot manipulators,
IEEE International Conference on Robotics and Automation, pp. 879-884
Dexter, A. L. (1995). Fuzzy Model-Based Fault Diagnosis, IEE Proceedings On Cont. Theory
and Applications, Vol. 142, No. 6, pp. 545-550
Dixon, W. E.; Walker, I. D.; Dawson, D. M. & Hartranft, J. P. (2000). Fault detection for robot
manipulators with parametric uncertainty: a prediction-error-based approach,
IEEE Transactions on Robotics and Automation, Vol. 16, No. 6, pp. 689-699
Fantuzzi, C.; Secchi, C.; Visioli A. (2003). On The Fault Detection And Isolation Of Industrial
Robot Manipulators, 7th International IFAC Symposium on Robot Control, pp. 399-404
Frank, P. M. & Ding, X. (1997). Survey of robust residual generation and evaluation methods
in observer-based fault detection systems, J. Proc. Cont., Vol.7, No.6, pp. 403-424.
Garcia, E. A. & Frank, P. M. (1997). Deterministic nonlinear observer-based Approaches To
Fault Diagnosis: A survey, Control Eng. Practice, Vol. 5, No. 5, pp. 663-670
Goel, P.; Dedeoglu, G.; Roumeliotis, S. I. & Sukhatme, G. S. (2000). Fault Detection and
Identification In A Mobile Robot Using Multiple Model Estimation and Neural
Network, IEEE International Conference on Robotics and Automation, pp. 2302-2309
Haykin, S. (1999). Neural Networks: A Compherensive Foundation, Prentice-Hall, 2. edition
Isermann, R.& Ballé, P. (1997). Trends in the application of model-based fault detection and
diagnosis of technical processes, Control Eng. Practice, Vol. 5, No. 5, pp. 709-719
Jang, J.R.; Sun, C.T. & Mizutani, E. (1997). Neuro-fuzzy And Soft Computing, Prentice-Hall Inc.
Kinnaert, M. (1999) Robust fault detection based on observers for bilinear systems,
Automatica, Vol. 35, pp. 1829-1842
Lee, I. S.; Kim, J. T.; Lee, J. W.; Lee, D. Y. & Kim K. Y. (2003). Model-based fault detection
and isolation method using ART2 neural network, International Journal Of Intelligent
Systems, Vol. 18, pp: 1087–1100
Leuschen, M. L.; Walker, I. D. & Cavallaro, J. R. (2005). Fault residual generation via
nonlinear analytical redundancy, IEEE Transactions on Control System Technology,
Vol. 13, No. 3, pp. 452-458

www.intechopen.com
152 Fault Detection

Lewis, F. L.; Abdallah, C. T. & Dawson D. M. (1993). Control Of Robot Manipulators,

MacMillan Publishing
Marcu, T.; Mirea, L. & Frank, P. M. (1998). Neural Observer Schemes For Robust Detection
And Isolation Of Process Faults, UKACC International Conference on CONTROL ‘98,
pp. 958-963
Moseler, O. & Isermann, R. (2000). Application of model based fault detection to a brushless
dc motor, IEEE Transactions on Industrial Electronics, Vol. 47, No. 5, pp. 1015-1020
Naughton, J. M.; Chen, Y.C. & Jiang J. (1996). A neural network application to fault
diagnosis for robotic manipulator, IEEE International Conference on Control
Applications, pp. 988-993
Patton, R. J.; Chen, J. & Lopez-Toribio, C. J. (1998). Fuzzy Observers for Nonlinear Dynamic
Systems Fault Diagnosis , 37th IEEE Conference on Decision & Control, pp. 84-89
Patton, R. J.; Frank, P. M. & Clark, R. N. (2000a). Issues Of Fault Diagnosis For Dynamic
Systems, Springer-Verlag
Patton, R.J.; Uppal, F. J.; Lopez-Toribio, C. J. (2000b). Soft Computing Approaches To Fault
Diagnosis For Dynamic Systems: A Survey, IFAC-SAFEPROCESS, pp. 298-311
Polycarpou, M. M. & Helmicki, A. J. (1995) Automated Fault Detection and
Accommodation: A Learning Systems Approach, IEEE Transactions On Systems,
Man and Cybernetics, Vol. 25, No. 11, pp. 1447-1458
Polycarpou, M. M. & Trunov, A.B. (2000). Learning Approach to Nonlinear Fault Diagnosis:
Detectability Analysis , IEEE Transactions On Automatic Control, Vol. 45, No. 4, pp.
806-812
Riedmiller, M. & Braun, H. (1993). A direct adaptive method for faster backpropagation
learning: the RPROP algorithm, IEEE International Conference on Neural Networks,
Vol. 1, pp. 586-591
Schneider, H. & Frank, P. M. (1996). Observer based supervision and fault detection in
robots using nonlinear and fuzzy logic residual evaluation, IEEE Transactions on
Control System Technology, Vol. 4, No. 3, pp. 274-282
Seliger, R. & Frank, P. M. (1991). Fault Diagnosis By Disturbance Decoupled Nonlinear
Observers, 30th Conf. on Decision and Control, pp. 2248-2253
Terra, M. H. & Tinós, R. (2001). Fault detection and isolation in robotic manipulators via
neural networks: a comparison among three architectures for residual analysis,
Journal of Robotic Systems, Vol. 18, No. 7, pp. 357-374
Tinós R. & Terra M. H. (2002). Free-Swinging and Locked Joint Fault Detection and Isolation
In Cooperative Manipulators , E. Symp. on Artificial Neural Networks, pp. 513-518
Vemuri, A. T. & Polycarpou, M. M. (1997). Neural-network-based robust fault diagnosis in
robotic systems, IEEE Transactions on Neural Networks, Vol. 8, No. 6, pp. 1410-1420
Venkatasubramanian, V.; Rengaswamy, R.; Kavuri, S.N. & Yin K. (2003). A review of
process fault detection and diagnosis part III: process history based methods,
Computers and Chemical Engineering, Vol.27, pp. 327-346
Yang, H. & Saif, M. (1995). State Observation, Failure Detection And Isolation(FDI) In
Bilinear Systems, 34th Conference on Decision & Control, pp. 2391-2396
Zhang, X.; Polycarpou, M. M. & Parisini, T. (2002). A Robust Detection and Isolation Scheme
for Abrupt and Incipient Faults in Nonlinear Systems, IEEE Transactions On
Automatic Control, Vol. 47, No. 4, pp. 576-592

www.intechopen.com
 Fault Detection
Edited by Wei Zhang

ISBN 978-953-307-037-7
Hard cover, 504 pages
Publisher InTech
Published online 01, March, 2010
Published in print edition March, 2010

In this book, a number of innovative fault diagnosis algorithms in recently years are introduced. These
methods can detect failures of various types of system effectively, and with a relatively high significance.

How to reference
In order to correctly reference this scholarly work, feel free to copy and paste the following:

Tolga Yuksel and Abdullah Sezgin (2010). Model-Based FDI Schemes For Robot Manipulators Using Soft
Computing Techniques, Fault Detection, Wei Zhang (Ed.), ISBN: 978-953-307-037-7, InTech, Available from:
http://www.intechopen.com/books/fault-detection/model-based-fdi-schemes-for-robot-manipulators-using-soft-
computing-techniques

InTech Europe InTech China

University Campus STeP Ri Unit 405, Office Block, Hotel Equatorial Shanghai
Slavka Krautzeka 83/A No.65, Yan An Road (West), Shanghai, 200040, China
51000 Rijeka, Croatia
Phone: +385 (51) 770 447 Phone: +86-21-62489820
Fax: +385 (51) 686 166 Fax: +86-21-62489821
www.intechopen.com