Académique Documents
Professionnel Documents
Culture Documents
AbstractThis paper presents a comprehensive review of the pable of application. Since the late 1980s, ANN started to be
industrial applications of artificial neural networks (ANNs), in utilized in a plethora of industrial applications.
the last 12 years. Common questions that arise to practitioners Nowadays, ANN are being applied to a lot of real world,
and control engineers while deciding how to use NNs for specific
industrial tasks are answered. Workable issues regarding imple- industrial problems, from functional prediction and system
mentation details, training and performance evaluation of such modeling (where physical processes are not well understood
algorithms are also discussed, based on a judiciously chronolog- or are highly complex), to pattern recognition engines and
ical organization of topologies and training methods effectively robust classifiers, with the ability to generalize while making
used in the past years. The most popular ANN topologies and decisions about imprecise input data. The ability of ANN to
training methods are listed and briefly discussed, as a reference
to the application engineer. Finally, ANN industrial applications learn and approximate relationships between input and output
are grouped and tabulated by their main functions and what is decoupled from the size and complexity of the problem
they actually performed on the referenced papers. The authors [49]. Actually, as relationships based on inputs and outputs
prepared this paper bearing in mind that an organized and are enriched, approximation capability improves. ANN offers
normalized review would be suitable to help industrial managing ideal solutions for speech, character and signal recognition.
and operational personnel decide which kind of ANN topology and
training method would be adequate for their specific problems. There are many different types of ANN. Some of the more
popular include multilayer perceptron (MLP) (which is generally
Index TermsArchitecture, industrial control, neural network trained with the back-propagation of error algorithm), learning
(NN) applications, training.
vector quantization, radial basis function (RBF), Hopfield and
Kohonen, to name a few. Some ANN are classified as feed
I. INTRODUCTION forward while others are recurrent (i.e., implement feedback)
depending on how data is processed through the network.
I N engineering and physics domains, algebraic and differen-
tial equations are used to describe the behavior and func-
tioning properties of real systems and to create mathematical
Another way of classifying ANN types is by their learning
method (or training), as some ANN employ supervised training,
models to represent them. Such approaches require accurate while others are referred to as unsupervised or self-organizing.
knowledge of the system dynamics and the use of estimation This paper concentrates on industrial applications of neural
techniques and numerical calculations to emulate the system networks (NNs). It was found that training methodology is more
operation. The complexity of the problem itself may introduce conveniently associated with a classification of how a certain
uncertainties, which can make the modeling nonrealistic or in- NN paradigm is supposed to be used for a particular indus-
accurate. Therefore, in practice, approximate analysis is used trial problem. There are some important questions to answer,
and linearity assumptions are usually made. Artificial neural in order to adopt an ANN solution for achieving accurate, con-
networks (ANNs) implement algorithms that attempt to achieve sistent and robust modeling. What is required to use an NN?
a neurological related performance, such as learning from ex- How are NNs superior to conventional methods? What kind of
perience, making generalizations from similar situations and problem functional characteristics should be considered for an
judging states where poor results were achieved in the past. ANN paradigm? What kind of structure and implementation
ANN history begins in the early 1940s. However, only in the should be used in accordance to an application in mind? This
mid-1980s these algorithms became scientifically sound and ca- article will follow a procedure that will bring such manage-
rial questions together and into a framework that can be used
to evaluate where and how such technology fits for industrial
Manuscript received October 23, 2001; revised September 20, 2002. Abstract applications, by laying out a classification scheme by means
published on the Internet March 4, 2003. This work was supported by the Na-
tional Science Foundation under Grant ECS 0134130.
of clustered concepts and distinctive characteristics of ANN
M. R. G. Meireles was with the Colorado School of Mines, Golden, CO engineering.
80401 USA. She is now with the Mathematics and Statistics Department, Pon-
tific Catholic University of Minas Gerais, 30.535-610 Belo Horizonte, Brazil
(e-mail: magali@pucminas.br). II. NN ENGINEERING
P. E. M. Almeida was with the Colorado School of Mines, Golden, CO 80401
USA. He is now with the Federal Center for Technological Education of Minas Before laying out the foundations for choosing the best ANN
Gerais, 30.510-000 Belo Horizonte, Brazil (e-mail: paulo@dppg.cefetmg.br). topology, learning method and data handling for classes of
M. G. Simes is with the Colorado School of Mines, Golden, CO 80401 USA
(e-mail: msimoes@mines.edu). industrial problems, it is important to understand how artificial
Digital Object Identifier 10.1109/TIE.2003.812470 intelligence (AI) evolved with required computational resources.
0278-0046/03$17.00 2003 IEEE
586 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 50, NO. 3, JUNE 2003
Artificial intelligence applications moved away from laboratory regarding the ANN topology selection are related to network
experiments to real world implementations. Therefore, soft- design, training, and practical considerations [25].
ware complexity also became an issue since conventional Von
Neumann machines are not suitable for symbolic processing,
B. Training Considerations
nondeterministic computations, dynamic execution, parallel,
distributed processing, and management of extensive knowl- Considerations, such as determining the input and output
edge bases [118]. variables, choosing the size of the training data set, initializing
In many AI applications, the knowledge needed to solve network weights, choosing training parameter values (such
a problem may be incomplete, because the source of the as learning rate and momentum rate), and selecting training
knowledge is unknown at the time the solution is devised, or stopping criteria, are important for several network topologies.
the environment may be changing and cannot be anticipated There is no generic formula that can be used to choose the
at design time. AI systems should be designed with an open parameter values. Some guidelines can be followed as an initial
concept that allows continuous refinement and acquisition of trial. After a few trials, the network designer should have
new knowledge. enough experience to set appropriate criteria that suit a given
There exist engineering problems for which finding the per- problem.
fect solution requires a practically impossible amount of re- The initial weights of an NN play a significant role in the con-
sources and an acceptable solution would be fine. NNs can give vergence of the training method. Without a priori information
good solutions for such classes of problems. Tackling the best about the final weights, it is a common practice to initialize all
ANN topology, learning method and data handling themselves weights randomly with small absolute values. In linear vector
become engineering approaches. The success of using ANN for quantization and derived techniques, it is usually required to
any application depends highly on the data processing, (i.e., data renormalize the weights at every training epoch. A critical pa-
handling before or during network operation). Once variables rameter is the speed of convergence, which is determined by
have been identified and data has been collected and is ready to the learning coefficient. In general, it is desirable to have fast
use, one can process it in several ways, to squeeze more infor- learning, but not so fast as to cause instability of learning iter-
mation out of and filter it. ations. Starting with a large learning coefficient and reducing
A common technique for coping with nonnormal data is to it as the learning process proceeds, results in both fast learning
perform a nonlinear transform to the data. To apply a transform, and stable iterations. The momentum coefficients are, usually,
one simply takes some function of the original variable and uses set according to a schedule similar to the one for the learning
the functional transform as a new input to the model. Commonly coefficients [128].
used nonlinear transforms include powers, roots, inverses, expo- Selection of training data plays a vital role in the performance
nentials, and logarithms [107]. of a supervised NN. The number of training examples used to
train an NN is sometimes critical to the success of the training
A. Assessment of NN Performance process. If the number of training examples is not sufficient,
then the network cannot correctly learn the actual inputoutput
An ANN must be used in problems exhibiting knottiness,
relation of the system. If the number of training examples is too
nonlinearity, and uncertainties that justify its utilization [45].
large, then the network training time will be longer. For some
They present the following features to cope with such complex-
applications, such as real-time adaptive neural control, training
ities:
time is a critical variable. For others, such as training the net-
learning from training data used for system identifi- work to perform fault detection, the training can be performed
cation; finding a set of connection strengths will allow off-line and more training data are preferred, over using insuffi-
the network to carry out the desired computation [96]; cient training data to achieve greater network accuracy. Gener-
generalization from inputs not previously presented ally, rather than focusing on volume, it is better to concentrate
during the training phase; by accepting an input and on the quality and representational nature of the data set. A good
producing a plausible response determined by the in- training set should contain routine, unusual and boundary-con-
ternal ANN connection structure makes such a system dition cases [8].
robust against noisy data, features exploited in indus- Popular criteria used to stop network training are small mean-
trial applications [59]; square training error and small changes in network weights.
mapping of nonlinearity making them suitable for Definition about how small is usually up to the network designer
identification in process control applications [90]; and is based on the desired accuracy level of the NN. Using
parallel processing capability, allowing fast pro- as an example a motor bearing fault diagnosis process, they
cessing for large-scale dynamical systems; used a learning rate of 0.01 and momentum of 0.8; the training
applicable to multivariable systems; they naturally was stopped when the root mean-square error of the training
process many inputs and have many outputs. set or the change in network weights was sufficiently small
used as a black-box approach (no prior knowledge for that application (less than 0.005) [72]. Therefore, if any
about a system) and implemented on compact proces- prior information about the relationship between inputs and
sors for space and power constrained applications. outputs is available and used correctly, the network structure
In order to select a good NN configuration, there are several and training time can be reduced and the network accuracy
factors to take into consideration. The major points of interest can be significantly improved.
MEIRELES et al.: INDUSTRIAL APPLICABILITY OF ANNs 587
model, to predict nationalism and social communications [120]. Topology of the networks: multilayered, single-lay-
However, his work remained almost unknown in the scien- ered, or recurrent. The network is multilayered if it has
tific community for more than ten years. In the early 1980s, distinct layers such as input, hidden and output. There
Hopfield introduced a recurrent-type NN, which was based are no connections among the neurons within the same
on Hebbian learning law. The model consisted of a set of layer. If each neuron can be connected with every other
first-order (nonlinear) differentiable equations that minimize a neuron in the network through directed edges, except
certain energy function [55]. In the mid-1980s, backpropaga- the output node, this network is called single layered
tion was rediscovered by two independent groups led by Parker (i.e., there is no hidden layer). A recurrent network
and Rumelhart et al., as the learning algorithm of feedforward distinguishes itself from the feedforward topologies in
NNs [88], [95]. Grossberg and Carpenter made significant con- that it has at least one feedback loop [49].
tributions with the Adaptive Resonance Theory (ART) in Data flow: recurrent or nonrecurrent. A nonrecur-
the mid-1980s, based on the idea that the brain spontaneously rent or feedforward model where the outputs always
organizes itself into recognition codes and neurons organize propagate from left to right in the diagrams. The
themselves to tune various and specific patterns defined as outputs from the input layer neurons propagate to the
self-organizing maps [20]. The dynamics of the network were right, becoming inputs to the hidden layer neurons and
modeled by first order differentiable equations based on im- then, outputs from the hidden layer neurons propagate
plementations of pattern clustering algorithms. Furthermore, to the right, becoming inputs to the output layer
Kosko extended some of the ideas of Grossberg and Hopfield
neurons. An NN, in which the outputs can propagate
to develop his adaptive Bi-directional Associative Memory
in both directions, forward and backward, is called a
(BAM) [67]. Hinton, Sejnowski, and Ackley developed the
recurrent model.
Boltzmann Machine, which is a kind of Hopfield net that
Types of input values: binary, bipolar or contin-
settles into solutions by a simulated annealing process as a
uous. Neurons in an artificial network can be defined to
stochastic technique [54]. Broomhead and Lowe first intro-
process different kinds of signals. The most common
duced RBF networks in 1988 [15]. Although the basic idea
types are binary (restricted to either 0 or 1), bipolar (ei-
of RBF was developed before, under the name method of po-
tential function, their work opened another NN frontier. Chen ther 1 or 1) and continuous (continuous real numbers
proposed functional-link networks (FLNs), where a nonlinear in a certain range).
functional transform of the network inputs aimed lower com- Forms of activation: linear, step, or sigmoid. Acti-
putational efforts and fast convergence [22]. vation functions will define the way neurons will be-
The 1988 DARPA NN study listed various NN applications have inside the network structure and, therefore, the
supporting the importance of such technology for commercial kind of relationship that will occur between input and
and industrial applications and triggering a lot of interest in output signals.
the scientific community, leading eventually to applications A common classification of ANNs is based on the way in
in industrial problems. Since then, the application of ANN to which their elements are interconnected. There is a study that
sophisticated systems has skyrocketed. NNs found widespread presents the approximate percentage of network utilization as:
relevance for several different fields. Our literature review MLP, 81.2%; Hopfield, 5.4%; Kohonen, 8.3%; and the others,
showed that practical industrial applications were reported in 5.1% [49]. This section will cover the main types of networks
peer-reviewed engineering journals from as early as 1988. that have been used in industrial applications, in a reasonable
Extensive use has been reported in pattern recognition and number of reports and applications. A comprehensive listing of
classification for image and speech recognition, optimization all available ANN structures and topologies is out of the scope
in planning of actions, motions, and tasks and modeling, of this discussion.
identification, and control.
Fig. 1 shows some industrial applications of NNs reported in A. MLPs
the last 12 years. The main purpose here is to give an idea of the In this structure, each neuron output is connected to every
most used ANN topologies and training algorithms and to relate neuron in subsequent layers connected in cascade with no con-
them to common fields in the industrial area. For each entry, nections between neurons in the same layer. A typical diagram
the type of the application, used ANN topology, implemented of this structure is detailed in Fig. 2.
training algorithm, and the main authors are presented. The col- MLP has been reported in several applications. Some ex-
lected data give a good picture of what has actually migrated amples are speed control of dc motors [94], [117], diagnostics
from academic research to practical industrial fields and shows
of induction motor faults [24], [25], [41], [42], induction
some of the authors and groups responsible for this migration.
motor control [17], [18], [56], [127], and current regulator for
pulsewidth-modulation (PWM) rectifiers [31]. Maintenance
IV. DILEMMATIC PIGEONHOLE OF NEURAL STRUCTURES and sensor failure detection was reported in [82], check valves
Choosing an ANN solution for an immediate application is operating in a nuclear power plant [57], [114], and vibration
a situation that requires a choice between options that are (or monitoring in rolling element bearings [2]. It was widely
seem) equally unfavorable or mutually exclusive. Several issues applied in feedback control [19], [40], [52], [59], [87], [89],
must be considered when regarding the problem point of view. [109], [110] and fault diagnosis of robotic systems [116].
The main features of an ANN can be classified as follows [81]. This structure was also used in a temperature control system
MEIRELES et al.: INDUSTRIAL APPLICABILITY OF ANNs 589
[63], [64], monitoring feed water flow rate and component fault diagnosis in a heat exchanger continuous stirred tank
thermal performance of pressurized water reactors [61], and reactor system [102]. It was used in a controller for turbo
590 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 50, NO. 3, JUNE 2003
F. RBF Networks
Fig. 8 shows the basic structure of an RBF network. The input
nodes pass the input values to the connecting arcs and the first
layer connections are not weighted. Thus, each hidden node
receives each input value unaltered. The hidden nodes are the
RBF units. The second layer of connections is weighted and the
output nodes are simple summations. This network does not ex-
tend to more layers.
For applications such as fault diagnosis, RBF networks offer
advantages over MLP. It is faster to train because training of
the two layers is decoupled [70]. This network was used to
improve the quantity and the quality of the galvanneale sheet Fig. 10. Polynomial ANN structure.
produced on galvanizing lines, integrating various approaches,
including quality monitoring, diagnosis, control, and optimiza- A practical advantage is that, unlike other networks, it operates
tion methods [13]. RBF was trained to evaluate and compare completely in parallel and the signal flows in a unique direction,
the different grasping alternatives by a robotic hand, according without a need for feedback from individual neurons to the
to geometrical and technological aspects of object surfaces, inputs. It can be used for mapping, classification and associative
as well as the specific task to be accomplished. The adoption memory, or to directly estimate a posteriori probabilities [103],
of RBF topology was justified by two reasons [34]. [104]. Probabilistic NNs were used to assist operators while
In most cases, it presents higher training speed identifying transients in nuclear power plants, such as plant
when compared with ANN based on back-propagation accident scenario, equipment failure or an external disturbance
training methods. to the system, at the earliest stages of their developments [6].
It allows an easier optimization of performance,
since the only parameter that can be used to modify its H. Polynomial Networks
structure is the number of neurons in the hidden layer. Fig. 10 depicts a polynomial network. It has its topology
Results using RBF networks were presented to illustrate that it is formed during the training process. Due to this feature, it is
possible to successfully control a generator system [43]. Power defined as a plastic network. The neuron activation function is
electronic drive controllers have also been implemented with based on elementary polynomials of arbitrary order. In this ex-
these networks, in digital signal processors (DSPs), to attain ro- ample, the network has seven inputs, although the network uses
bust properties [38]. only five of them. This is due to the automatic input selection
capability of the training algorithm. Automatic feature selection
G. Probabilistic NNs (PNNs) is very useful in control applications when the plant model order
PNNs are somewhat similar in structure to MLPs. Some is unknown. Each neuron output can be expressed by a second-
basic differences among them are the use of activation by order polynomial function
exponential functions and the connection patterns between , where and are inputs and
neurons. As a matter of fact, the neurons at the internal layers and are polynomial coefficients which are equivalent to the
are not fully connected, depending on the application in turn. network weights and is the neuron output. The Group Method
Fig. 9 depicts this structure, showing its basic differences from of Data Handling (GMDH) is a statistics-based training method
ordinary MLP structure. PNN training is normally easy and largely used in modeling economic, ecological, environmental
instantaneous, because of the smaller number of connections. and medical problems. The GMDH training algorithm can be
MEIRELES et al.: INDUSTRIAL APPLICABILITY OF ANNs 593
where
(10)
(11)
The idea was to reduce the possibility of the network being C. Second-Order Gradient Methods
trapped in the local minimum. These methods use the Hessian matrix , which is the matrix
DeltaBarDelta (DBD): The DBD learning rules use adap- of second derivatives of with respect to the weights . This
tive learning rates, to speed up the convergence. The adaptive matrix contains information about how the gradient changes in
learning rate adopted is based on a local optimization method. different directions in weight space
This technique uses gradient descent for the search direction and
then applies individual step sizes for each weight. This means (12)
the actual direction taken in weight space is not necessarily
along the line of the steepest gradient. If the weight updates be-
Newton Method: The Newton method weights update is pro-
tween consecutive iterations are in opposite directions, the step
size is decreased; otherwise, it is increased. This is prompted by cessed as follows:
the idea that if the weight changes are oscillating, the minimum
(13)
is between the oscillations and a smaller step size might find
that minimum. The step size may be increased again once the
error has stopped oscillating. However, the Newton method is not commonly used because
If denotes the learning rate for the weight , then computing the Hessian matrix is computationally expensive.
Furthermore, the Hessian matrix may not be positive definite
at every point in the error surface. To overcome the problem,
(8)
several methods have being proposed to approximate the Hes-
sian matrix [84].
and is as follows: GaussNewton Method: The GaussNewton method pro-
duces an matrix that is an approximation to the Hessian
matrix, having elements represented by
(9)
otherwise (14)
596 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 50, NO. 3, JUNE 2003
where . However, the eligibility traces can be generated using the following linear
GaussNewton method may still have ill conditioning if difference equation:
close to or is singular [75].
LevenbergMarquardt (LM) Method: The LM method over- (22)
comes this difficulty by including an additional term, which
is added to the GaussNewton approximation of the Hessian where determines the trace decay rate, is the
giving input, and is the output.
Adaptive Critic Learning: The weights update in a critic net-
(15) work is as follows [9]:
Kohonen Self-Organizing Learning: The network is trained varied until satisfactory convergence is achieved. Currently,
according to the following algorithm, frequently called the the most widely used algorithm for training MLPs is the back-
winner-takes-all rule. propagation algorithm. It minimizes the mean square error
1) Apply an input vector . between the desired and the actual output of the network. The
2) Calculate the distance (in -dimensional space) be- optimization is carried out with a gradient-descent technique.
tween and the weight vectors of each unit. In Eu- There are two critical issues in network learning: estimation
clidean space, this is calculated as follows: error and training time. These issues may be affected by the
network architecture and the training set. The network archi-
(28) tecture includes the number of hidden nodes, number of hidden
3) The unit that has the weight vector closest to is declared layers and values of learning parameters. The training set is
the winner unit. This weight vector, called , becomes related to the number of training patterns, inaccuracies of input
the center of a group of weight vectors that lie within a data and preprocessing of data. The backpropagation algorithm
distance from . does not always find the global minimum, but may stop at a
4) For all weight vectors within a distance of , train local minimum. In practice, the type of minimum has little
this group of nearby weight vectors according to the for- importance, as long as the desired mapping or classification
mula that follows: is reached with a desired accuracy. The optimization criterion
of the backpropagation algorithm is not very good, from the
(29) pattern recognition point of view. The algorithm minimizes
the square error between the actual and the desired output, not
5) Perform steps 1)4), cycling through each input vector
the number of faulty classifications, which is the main goal
until convergence.
in pattern recognition. The algorithm is too slow for prac-
F. Practical Considerations tical applications, especially if many hidden layers are used.
In addition, a backpropagation net has poor memory. When
NNs are unsurpassed at identifying patterns or trends in data the net learns something new it forgets the old. Despite its
and well suited for prediction or forecasting needs including shortcomings, bac-propagation is broadly used. Although the
sales and customer research, data validation, risk management, back-propagation algorithm has been a significant milestone,
and industrial process control. One of the fascinating aspects, of many attempts have been made to speed up the convergence
the practical implementation of NNs to industrial applications, and significant improvement are observed by using various
is the ability to manage data interaction between electrical and second order approaches, namely, Newtons method, conjugate
mechanical behavior and often other disciplines, as well. The gradients, or the LM optimization technique [5], [10], [21],
majority of the reported applications involve fault diagnosis and [48]. The issues to be dealt with are [84], [102] as follows:
detection, quality control, pattern recognition, and adaptive con-
trol [14], [44], [74], [115]. Supervised NNs can mimic the be- 1) slow convergence speed;
havior of human control systems, as long as data corresponding 2) sensitivity to initial conditions;
to the human operator and the control input are supplied [7], 3) trapping in local minima;
[126]. Most of the existing, successful applications in control 4) instability if learning rate is too large.
use supervised learning, or any form of a reinforcement learning One of the alternatives for the problem of being trapped in a local
approach that is also supervised. Unsupervised learning is not minimum is adding the momentum term using the BPM, which
suitable, particularly for online control, due the slow adaptation also improves the convergence speed. Another alternative used
and required time for the network to settle into stable condi- when a backpropagation algorithm is difficult to implement,
tions. Unsupervised learning schemes are used mostly for pat- as in analog hardware, is the random weight change (RWC).
tern recognition, by defining group of patterns into a number of This algorithm has shown to be immune to offset sensitivity
clusters or classes. and nonlinearity errors. It is a stochastic learning that makes
There are some advantages to NNs over multiple data re- sure that the error function decreases on average, since it is
gression. There is no need to select the most important inde- going up or down at any one time. It is often called simulated
pendent variables in the data set. The synapses associated with annealing because of its operational similarity to annealing
irrelevant variables readily show negligible weight values; in processes [73]. Second order gradient methods use the matrix
their turn, relevant variables present significant synapse weight of second derivatives of with respect to the weights .
values. There is also no need to propose a model function However, computing this matrix is computationally expensive
as required in multiple regressions. The learning capability of and the methods presented tried to approximate this matrix
NNs allows them to discover more complex and subtle in- to make algorithms more accessible.
teractions between the independent variables, contributing to Linear rewardpenalty, associative search, and adaptive critic
the development of a model with maximum precision. NNs algorithms are characterized as a special case of supervised
are intrinsically robust showing more immunity to noise, an learning called reinforcement learning. They do not need to
important factor in modeling industrial processes. NNs have explicitly compute derivatives. Computation of derivatives
been applied within industrial domains, to address the inherent usually introduces a lot of high frequency noise in the control
complexity of interacting processes under the lack of robust loop. Therefore, they are very suitable for some complex
analytical models of real industrial processes. In many cases, systems, where basic training algorithms may fail or produce
network topologies and training parameters are systematically suboptimal results. On the other hand, those methods present
598 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 50, NO. 3, JUNE 2003
slower learning processes; and, because of this, they are adopted themselves, have to be found from data in the training phase,
especially in the cases where only a single bit of information without supervision.
(for example, whether the output is right or wrong) is available.
C. Process Control
VI. TAXONOMY OF NN APPLICATIONS The NN makes use of nonlinearity, learning, parallel pro-
From the viewpoint of industrial applications, ANN applica- cessing, and generalization capabilities for application to
tions can be divided into four main categories. advanced intelligent control. They can be classified into some
major methods, such as supervised control, inverse control,
A. Modeling and Identification neural adaptive control, back-propagation of utility (which is
an extended method of a back-propagation through time) and
Modeling and identification are techniques to describe the adaptive critics (which is an extended method of reinforcement
physical principles existing between the input and the output learning algorithm) [2]. MLP structures were used for digital
of a system. The ANN can approximate these relationships current regulation of inverter drives [16], to predict trajectories
independent of the size and complexity of the problem. It has in robotic environments [19], [40], [52], [73], [79], [87], [89],
been found to be an effective system for learning discriminants, [110], to control turbo generators [117], to monitor feed water
for patterns from a body of examples. MLP is used as the flow rate and component thermal performance of pressurized
basic structure for a bunch of applications [4], [12], [17], [18], water reactors [61], to regulate temperature [64], and to predict
[32], [46], [56], [85], [94], [119], [127]. The Hopfield network natural gas consumption [65]. Dynamical versions of MLP
can be used to identify problems of linear time-varying or networks were used to control a nonlinear dynamic model of
time-invariant systems [28]. Recurrent network topology [36], a robot [60], [97], to control manufacturing cells [92], and to
[83] has received considerable attention for the identification implement a programmable cascaded low-pass filter [101]. A
of nonlinear dynamical systems. A functional-link NN ap- dynamic MLP is a classical MLP structure where the outputs
proach (FLN) was used to perform thermal dynamical system are fed back to the inputs by means of time delay elements.
identification [113]. Other structures can be found as functional link networks to
control robots [69] and RBF networks to predict, from operating
B. Optimization and Classification
conditions and from features of a steel sheet, the thermal en-
Optimization is often required for design, planning of ac- ergy required to correct alloying [13]. RBF networks can be ob-
tions, motions, and tasks. However, as known in the Traveling served, as well, in predictive controllers for drive systems [38].
Salesman Problem, many parameters cause the amount of cal- Hopfield structures were used for torque minimization control
culation to be tremendous and the ordinary method cannot be of redundant manipulators [33]. FPNs can be used for function
applied. An affective approach is to find the optimal solution approximation inside specific control schemes [100]. CMAC
by defining an energy function and using the NN with parallel networks were implemented in research automobiles [76] and
processing, learning and self-organizing capabilities to operate to control robots [79].
in such a way that the energy is reduced. It is shown that ap-
plication of the optimal approach makes effective use of ANN D. Pattern Recognition
sensing, recognizing, and forecasting capabilities, in the control Some specific ANN structures, such as Kohonen and proba-
of robotic manipulators with impact taken into account [45]. bilistic networks, are studied and applied mainly for image and
Classification using an ANN can also be viewed as an voice recognition. Research in image recognition includes ini-
optimization problem, provided that the existing rules to dis- tial vision (stereo vision of both eyes, outline extraction, etc.)
tinguish the various classes of events/materials/objects can be close to the biological (particularly brain) function, manually
described in functional form. In such cases, the networks will written character recognition by cognition at the practical level
decide if a particular input belongs to one of the defined classes and cell recognition for mammalian cell cultivation by using
by optimizing the functional rules and a posteriori evaluating NNs [45]. Kohonen networks were used for image inspection
the achieved results. Different authors [34], [70] have proposed and for disease identification from mammographic images [98].
RBF approaches. For applications such as fault diagnosis, RBF Probabilistic networks were used for transient detection to en-
networks offer clear advantages over MLPs. They are faster hance nuclear reactors operational safety [6]. As in the others
to train, because layer training is decoupled [70]. Cellular categories, the MLP is widely used as well. The papermaking
networks [47], ART networks [122], and Hopfield networks industry [39] is one such example.
[105], [112] can be used as methods to detect, to isolate faults,
and to promote industrial quality control. MLPs are also widely
VII. CONCLUSION
used for these purposes [23], [25], [35], [102], [114], [121].
This structure can be found in induction motor [41], [42] and This paper has described theoretical aspects of NNs related to
bearing [72] fault diagnosis, for nondestructive evaluation of their relevance for industrial applications. Common questions
check valve performance and degradation [2], [57], in defect that an engineer would ask when choosing an NN for a par-
detection on woven fabrics [99], and in robotic systems [116]. ticular application were answered. Characteristics of industrial
Finally, it is important to mention clustering applications, processes, which would justify the ANN utilization, were dis-
which are special cases of classification where there is no cussed and some areas of importance were proposed. Important
supervision during the training phase. The relationships be- structures and training methods, with relevant references that il-
tween elements of the existing classes and even the classes lustrated the utilization of those concepts, were presented.
MEIRELES et al.: INDUSTRIAL APPLICABILITY OF ANNs 599
This survey observed that, although ANNs have a history [19] R. Carelli, E. F. Camacho, and D. Patio, A neural network based feed
of more than 50 years, most of industrial applications were forward adaptive controller for robots, IEEE Trans. Syst., Man, Cy-
bern., vol. 25, pp. 12811288, Sept. 1995.
launched in the last ten years, where it was justified that the [20] G. A. Carpenter and S. Grossberg, Associative learning, adaptive
investigators provided either an alternative or a complement to pattern recognition and cooperative- competitive decision making, in
other classical techniques. Those ANN applications demon- Optical and Hybrid Computing, H. Szu, Ed. Bellingham, WA: SPIE,
1987, vol. 634, pp. 218247.
strated adaptability features integrated with the industrial [21] C. Charalambous, Conjugate gradient algorithm for efficient training
problem, thus becoming part of the industrial processes. The of artificial neural networks, Proc. Inst. Elect. Eng., vol. 139, no. 3, pp.
authors firmly believe that such an intricate field of NNs is just 301310, 1992.
[22] S. Chen and S. A. Billings, Neural networks for nonlinear dynamic
starting to permeate a broad range of interdisciplinary problem system modeling and identification, Int. J. Control, vol. 56, no. 2, pp.
solving streams. The potential of NNs will be integrated into a 319346, 1992.
still larger and all-encompassing field of intelligence systems [23] R. P. Cherian, L. N. Smith, and P. S. Midha, A neural network approach
for selection of powder metallurgy materials and process parameters,
and will soon be taught for students and engineers as an Artif. Intell. Eng., vol. 14, pp. 3944, 2000.
ordinary mathematical tool. [24] M. Y. Chow, P. M. Mangum, and S. O. Yee, A neural network approach
to real-time condition monitoring of induction motors, IEEE Trans. Ind.
Electron., vol. 38, pp. 448453, Dec. 1991.
REFERENCES [25] M. Y. Chow, R. N. Sharpe, and J. C. Hung, On the application and
design of artificial neural networks for motor fault detectionPart II,
[1] J. S. Albus, A new approach to manipulator control: The cerebellar IEEE Trans. Ind. Electron., vol. 40, pp. 189196, Apr. 1993.
model articulation controller, Trans. ASME, J. Dyn. Syst., Meas. Con- [26] , On the application and design of artificial neural networks for
trol, vol. 97, pp. 220227, Sept. 1975. motor fault detectionPart I, IEEE Trans. Ind. Electron., vol. 40, pp.
[2] I. E. Algundigue and R. E. Uhrig, Automatic fault recognition in me- 181188, Apr. 1993.
chanical components using coupled artificial neural networks, in Proc. [27] T. W. S. Chow and Y. Fang, A recurrent neural-network based real-time
IEEE World Congr. Computational Intelligence, JuneJuly 1994, pp. learning control strategy applying to nonlinear systems with unknown
33123317. dynamics, IEEE Trans. Ind. Electron., vol. 45, pp. 151161, Feb. 1998.
[3] P. E. M. Almeida and M. G. Simes, Fundamentals of a fast conver- [28] S. R. Chu and R. Shoureshi, Applications of neural networks in learning
gence parametric CMAC network, in Proc. IJCNN01, vol. 3, 2001, of dynamical systems, IEEE Trans. Syst., Man, Cybern., vol. 22, pp.
pp. 30153020. 160164, Jan./Feb. 1992.
[4] K. Andersen, G. E. Cook, G. Karsai, and K. Ramaswamy, Artificial [29] S. R. Chu, R. Shoureshi, and M. Tenorio, Neural networks for system
neural networks applied to arc welding process modeling and control, identification, IEEE Contr. Syst. Mag., vol. 10, pp. 3135, Apr. 1990.
IEEE Trans. Ind. Applicat., vol. 26, pp. 824830, Sept./Oct. 1990. [30] L. O. Chua, T. Roska, T. Kozek, and . Zarndy, The CNN Para-
[5] T. J. Andersen and B. M. Wilamowski, A modified regression algorithm digmA short tutorial, in Cellular Neural Networks, T. Roska and J.
for fast one layer neural network training, in Proc. World Congr. Neural Vandewalle, Eds. New York: Wiley, 1993, pp. 114.
Networks, vol. 1, Washington DC, July 1721, 1995, pp. 687690. [31] M. Cichowlas, D. Sobczuk, M. P. Kazmierkowski, and M. Malinowski,
[6] I. K. Attieh, A. V. Gribok, J. W. Hines, and R. E. Uhrig, Pattern recogni- Novel artificial neural network based current controller for PWM rec-
tion techniques for transient detection to enhance nuclear reactors oper- tifiers, in Proc. 9th Int. Conf. Power Electronics and Motion Control,
ational safety, in Proc. 25th CNS/CNA Annu. Student Conf., Knoxville, 2000, pp. 4146.
TN, Mar. 2000. [32] G. E. Cook, R. J. Barnett, K. Andersen, and A. M. Strauss, Weld mod-
[7] S. M. Ayala, G. Botura Jr., and O. A. Maldonado, AI automates substa- eling and control using artificial neural network, IEEE Trans. Ind. Ap-
tion control, IEEE Comput. Applicat. Power, vol. 15, pp. 4146, Jan. plicat., vol. 31, pp. 14841491, Nov./Dec. 1995.
2002. [33] H. Ding and S. K. Tso, A fully neural-network-based planning scheme
[8] D. L. Bailey and D. M. Thompson, Developing neural-network appli- for torque minimization of redundant manipulators, IEEE Trans. Ind.
cations, AI Expert, vol. 5, no. 9, pp. 3441, 1990. Electron., vol. 46, pp. 199206, Feb. 1999.
[9] A. G. Barto, R. S. Sutton, and C. W. Anderson, Neuronlike elements [34] G. Dini and F. Failli, Planning grasps for industrial robotized applica-
that can solve difficult control problems, IEEE Trans. Syst., Man, Cy- tions using neural networks, Robot. Comput. Integr. Manuf., vol. 16,
bern., vol. SMC-13, pp. 834846, Sept./Oct. 1983. pp. 451463, Dec. 2000.
[10] R. Battiti, First- and second-order methods for learning: Between [35] M. Dolen and R. D. Lorenz, General methodologies for neural net-
steepest descent and Newtons method, Neural Computation, vol. 4, work programming, in Proc. IEEE Applied Neural Networks Conf.,
no. 2, pp. 141166, 1992. Nov. 1999, pp. 337342.
[11] B. Bavarian, Introduction to neural networks for intelligent control, [36] , Recurrent neural network topologies for spectral state estimation
IEEE Contr. Syst. Mag., vol. 8, pp. 37, Apr. 1988. and differentiation, in Proc. ANNIE Conf., St. Louis, MO, Nov. 2000.
[12] N. V. Bhat, P. A. Minderman, T. McAvoy, and N. S. Wang, Modeling [37] M. Dolen, P. Y. Chung, E. Kayikci, and R. D. Lorenz, Disturbance
chemical process systems via neural computation, IEEE Contr. Syst. force estimation for CNC machine tool feed drives by structured neural
Mag., vol. 10, pp. 2430, Apr. 1990. network topologies, in Proc. ANNIE Conference, St. Louis, MO, Nov.
[13] G. Bloch, F. Sirou, V. Eustache, and P. Fatrez, Neural intelligent control 2000.
for a steel plant, IEEE Trans. Neural Networks, vol. 8, pp. 910918, [38] Y. Dote, M. Strefezza, and A. Suyitno, Neuro fuzzy robust controllers
July 1997. for drive systems, in Proc. IEEE Int. Symp. Industrial Electronics,
[14] Z. Boger, Experience in developing models of industrial plants by large 1993, pp. 229242.
scale artificial neural networks, in Proc. Second New Zealand Interna- [39] P. J. Edwards, A. F. Murray, G. Papadopoulos, A. R. Wallace, J. Barnard,
tional Two-Stream Conf. Artificial Neural Networks and Expert Systems, and G. Smith, The application of neural networks to the papermaking
1995, pp. 326329. industry, IEEE Trans. Neural Networks, vol. 10, pp. 14561464, Nov.
[15] D. S. Broomhead and D. Lowe, Multivariable functional interpolation 1999.
and adaptive network, Complex Syst., vol. 2, pp. 321355, 1988. [40] M. J. Er and K. C. Liew, Control of adept one SCARA robot using
[16] M. Buhl and R. D. Lorenz, Design and implementation of neural net- neural networks, IEEE Trans. Ind. Electron., vol. 44, pp. 762768, Dec.
works for digital current regulation of inverter drives, in Conf. Rec. 1997.
IEEE-IAS Annu. Meeting, 1991, pp. 415423. [41] F. Filippetti, G. Franceschini, and C. Tassoni, Neural networks aided
[17] B. Burton and R. G. Harley, Reducing the computational demands of on-line diagnostics of induction motor rotor faults, IEEE Trans. Ind.
continually online-trained artificial neural networks for system identi- Applicat., vol. 31, pp. 892899, July/Aug. 1995.
fication and control of fast processes, IEEE Trans. Ind. Applicat., vol. [42] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, Recent develop-
34, pp. 589596, May/June 1998. ments of induction motor drives fault diagnosis using AI techniques,
[18] B. Burton, F. Kamran, R. G. Harley, T. G. Habetler, M. Brooke, and R. IEEE Trans. Ind. Electron., vol. 47, pp. 9941004, Oct. 2000.
Poddar, Identification and control of induction motor stator currents [43] D. Flynn, S. McLoone, G. W. Irwin, M. D. Brown, E. Swidenbank, and
using fast on-line random training of a neural network, in Conf. Rec. B. W. Hogg, Neural control of turbogenerator systems, Automatica,
IEEE-IAS Annu. Meeting, 1995, pp. 17811787. vol. 33, no. 11, pp. 19611973, 1997.
600 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 50, NO. 3, JUNE 2003
[44] D. B. Fogel, Selecting an optimal neural network industrial electronics [71] F. L. Lewis, A. Yesildirek, and K. Liu, Multilayer neural-net robot con-
society, in Proc. IEEE IECON90, vol. 2, 1990, pp. 12111214. troller with guaranteed tracking performance, IEEE Trans. Neural Net-
[45] T. Fukuda and T. Shibata, Theory and applications of neural networks works, vol. 7, pp. 388399, Mar. 1996.
for industrial control systems, IEEE Trans. Ind. Applicat., vol. 39, pp. [72] B. Li, M. Y. Chow, Y. Tipsuwan, and J. C. Hung, Neural-network based
472489, Nov./Dec. 1992. motor rolling bearing fault diagnosis, IEEE Trans. Ind. Electron., vol.
[46] A. A. Gorni, The application of neural networks in the modeling of 47, pp. 10601069, Oct. 2000.
plate rolling processes, JOM-e, vol. 49, no. 4, electronic document, Apr. [73] J. Liu, B. Burton, F. Kamran, M. A. Brooke, R. G. Harley, and T. G.
1997. Habetler, High speed on-line neural network of an induction motor im-
[47] N. Guglielmi, R. Guerrieri, and G. Baccarani, Highly constrained mune to analog circuit nonidealities, in Proc. IEEE Int. Symp. Circuits
neural networks for industrial quality control, IEEE Trans. Neural and Systems, June 1997, pp. 633636.
Networks, vol. 7, pp. 206213, Jan. 1996. [74] Y. Liu, B. R. Upadhyaya, and M. Naghedolfeizi, Chemometric data
[48] M. T. Hagan and M. Menhaj, Training feedforward networks with analysis using artificial neural networks, Appl. Spectrosc., vol. 47, no.
the Marquardt algorithm, IEEE Trans. Neural Networks, vol. 5, pp. 1, pp. 1223, 1993.
989993, Nov. 1994. [75] L. Ljung, System Identification: Theory for the User. New York: Pren-
[49] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd tice-Hall, 1987.
ed. New York: Prentice-Hall, 1995. [76] M. Majors, J. Stori, and D. Cho, Neural network control of automotive
[50] S. M. Halpin and R. F. Burch, Applicability of neural networks to indus- fuel-injection systems, IEEE Contr. Syst. Mag., vol. 14, pp. 3136, June
trial and commercial power systems: A tutorial overview, IEEE Trans. 1994.
Ind. Applicat., vol. 33, pp. 13551361, Sept./Oct. 1997. [77] W. S. McCulloch and W. Pitts, A logical calculus of the ideas immanent
[51] H. Harrer and J. Nossek, Discrete-time cellular neural networks, in in nervous activity, Bull. Math. Biophys., vol. 5, pp. 115133, 1943.
Cellular Neural Networks, T. Roska and J. Vandewalle, Eds. New [78] M. Milanova, P. E. M. Almeida, J. Okamoto Jr., and M. G. Simes, Ap-
York: Wiley, 1993, pp. 1529. plications of cellular neural networks for shape from shading problem,
[52] H. Hashimoto, T. Kubota, M. Sato, and F. Harashima, Visual control of in Proc. Int. Workshop Machine Learning and Data Mining in Pattern
robotic manipulator based on neural networks, IEEE Trans. Ind. Elec- Recognition, Lecture Notes in Artificial Intelligence, P. Perner and M.
tron., vol. 39, pp. 490496, Dec. 1992. Petrou, Eds., Leipzig, Germany, September 1999, pp. 5263.
[53] D. O. Hebb, The Organization of Behavior. New York: Wiley, 1949. [79] W. T. Miller III, Real-time application of neural networks for sensor-
[54] G. E. Hinton and T. J. Sejnowski, Learning and relearning in Boltzmann based control of robots with vision, IEEE Trans. Syst., Man, Cybern.,
machines, in The PDP Research Group, D. Rumelhart and J. McClel- vol. 19, pp. 825831, July/Aug. 1989.
land, Eds. Cambridge, MA: MIT Press, 1986. [80] M. L. Minsky and S. Papert, Perceptrons: An Introduction to Computa-
[55] J. J. Hopfield, Neural networks and physical systems with emergent tional Geometry. Cambridge, MA: MIT Press, 1969.
collective computational abilities, in Proc. Nat. Acad. Sci., vol. 79, Apr. [81] T. Munakata, Fundamentals of the New Artificial IntelligenceBeyond
1982, pp. 24452558. Traditional Paradigms. Berlin, Germany: Springer-Verlag, 1998.
[56] C. Y. Huang, T. C. Chen, and C. L. Huang, Robust control of induction [82] S. R. Naidu, E. Zafiriou, and T. J. McAvoy, Use of neural networks for
motor with a neural-network load torque estimator and a neural-network sensor failure detection in a control system, IEEE Contr. Syst. Mag.,
identification, IEEE Trans. Ind. Electron., vol. 46, pp. 990998, Oct. vol. 10, pp. 4955, Apr. 1990.
1999. [83] K. S. Narendra and K. Parthasarathy, Identification and control of dy-
[57] A. Ikonomopoulos, R. E. Uhrig, and L. H. Tsoukalas, Use of neural namical systems using neural networks, IEEE Trans. Neural Networks,
networks to monitor power plant components, in Proc. American Power vol. 1, pp. 427, Mar. 1990.
Conf., vol. 54-II, Apr. 1992, pp. 11321137. [84] G. W. Ng, Application of Neural Networks to Adaptive Control of Non-
[58] M. Jelnek. (1999) Everything you wanted to know about ART linear Systems. London, U.K.: Research Studies Press, 1997.
neural networks, but were afraid to ask. [Online]. Available: [85] D. H. Nguyen and B. Widrow, Neural networks for self-learning control
http://cs.felk.cvut.cz/~xjeline1/semestralky/nan systems, IEEE Contr. Syst. Mag., vol. 10, pp. 1823, Apr. 1990.
[59] S. Jung and T. C. Hsia, Neural network impedance force control of [86] J. R. Noriega and H. Wang, A direct adaptive neural-network control
robot manipulator, IEEE Trans. Ind. Electron., vol. 45, pp. 451461, for unknown nonlinear systems and its application, IEEE Trans. Neural
June 1998. Networks, vol. 9, pp. 2734, Jan. 1998.
[60] A. Karakasoglu and M. K. Sundareshan, A recurrent neural network- [87] T. Ozaki, T. Suzuki, T. Furuhashi, S. Okuma, and Y. Uchikawa, Tra-
based adaptive variable structure model-following control of robotic ma- jectory control of robotic manipulators using neural networks, IEEE
nipulators, Automatica, vol. 31, no. 10, pp. 14951507, 1995. Trans. Ind. Electron., vol. 38, June 1991.
[61] K. Kavaklioglu and B. R. Upadhyaya, Monitoring feedwater flow [88] D. B. Parker, A comparison of algorithms for neuron-like cells, in
rate and component thermal performance of pressurized water reactors Neural Networks for Computing, J. S. Denker, Ed. New York: Amer-
by means of artificial neural networks, Nucl. Technol., vol. 107, pp. ican Institute of Physics, 1986, pp. 327332.
112123, July 1994. [89] P. Payeur, H. Le-Huy, and C. M. Gosselin, Trajectory prediction for
[62] M. Kawato, Y. Uno, M. Isobe, and R. Suzuki, Hierarchical neural net- moving objects using artificial neural networks, IEEE Trans. Ind. Elec-
work model for voluntary movement with application to robotics, IEEE tron., vol. 42, pp. 147158, Apr. 1995.
Contr. Syst. Mag., vol. 8, pp. 815, Apr. 1988. [90] M. H. Rahman, R. Fazlur, R. Devanathan, and Z. Kuanyi, Neural net-
[63] M. Khalid and S. Omatu, A neural network controller for a temperature work approach for linearizing control of nonlinear process plants, IEEE
control system, IEEE Contr. Syst. Mag., vol. 12, pp. 5864, June 1992. Trans. Ind. Electron., vol. 47, pp. 470477, Apr. 2000.
[64] M. Khalid, S. Omatu, and R. Yusof, Temperature regulation with neural [91] F. Rosenblatt, The perceptron: A probabilistic model for information
networks and alternative control schemes, IEEE Trans. Neural Net- storage and organization in the brain, Psych. Rev., vol. 65, pp. 386408,
works, vol. 6, pp. 572582, May 1995. 1958.
[65] A. Khotanzad, H. Elragal, and T. L. Lu, Combination of artificial [92] G. A. Rovithakis, V. I. Gaganis, S. E. Perrakis, and M. A. Christodoulou,
neural-network forecasters for prediction of natural gas consumption, Real-time control of manufacturing cells using dynamic neural net-
IEEE Trans. Neural Networks, vol. 11, pp. 464473, Mar. 2000. works, Automatica, vol. 35, no. 1, pp. 139149, 1999.
[66] Y. H. Kim, F. L. Lewis, and D. M. Dawson, Intelligent optimal control [93] A. Rubaai and M. D. Kankam, Adaptive real-time tracking controller
of robotic manipulators using neural network, Automatica, vol. 36, no. for induction motor drives using neural designs, in Conf. Rec. IEEE-IAS
9, pp. 13551364, 2000. Annu. Meeting, vol. 3, Oct. 1996, pp. 17091717.
[67] B. Kosko, Adaptive bi-directional associative memories, Appl. Opt., [94] A. Rubaai and R. Kotaru, Online identification and control of a DC
vol. 26, pp. 49474960, 1987. motor using learning adaptation of neural networks, IEEE Trans. Ind.
[68] S. Y. Kung and J. N. Hwang, Neural network architectures for robotic Applicat., vol. 36, pp. 935942, May/June 2000.
applications, IEEE Trans. Robot. Automat., vol. 5, pp. 641657, Oct. [95] D. D. Rumelhart, G. E. Hinton, and R. J. Williams, Learning repre-
1989. sentations by back-propagating errors, Nature, vol. 323, pp. 533536,
[69] C. Kwan, F. L. Lewis, and D. M. Dawson, Robust neural-network con- 1986.
trol of rigid-link electrically driven robots, IEEE Trans. Neural Net- [96] D. E. Rumelhart, B. Widrow, and M. A. Lehr, The basic ideas in neural
works, vol. 9, pp. 581588, July 1998. networks, Commun. ACM, vol. 37, no. 3, pp. 8792, Mar. 1994.
[70] J. A. Leonard and M. A. Kramer, Radial basis function networks for [97] M. Saad, P. Bigras, L. A. Dessaint, and K. A. Haddad, Adaptive robot
classifying process faults, IEEE Contr. Syst. Mag., vol. 11, pp. 3138, control using neural networks, IEEE Trans. Ind. Electron., vol. 41, pp.
Apr. 1991. 173181, Apr. 1994.
MEIRELES et al.: INDUSTRIAL APPLICABILITY OF ANNs 601
[98] S. Sardy and L. Ibrahim, Experimental medical and industrial applica- [122] J. R. Whiteley, J. F. Davis, A. Mehrotra, and S. C. Ahalt, Observations
tions of neural networks to image inspection using an inexpensive per- and problems applying ART2 for dynamic sensor pattern interpretation,
sonal computer, Opt. Eng., vol. 35, no. 8, pp. 21822187, Aug. 1996. IEEE Trans. Syst., Man, Cybern. A, vol. 26, pp. 423437, July 1996.
[99] S. Sardy, L. Ibrahim, and Y. Yasuda, An application of vision system [123] B. Widrow, DARPA Neural Network Study. Fairfax, VA: Armed Forces
for the identification and defect detection on woven fabrics by using Communications and Electronics Assoc. Int. Press, 1988.
artificial neural networks, in Proc. Int. Joint Conf. Neural Networks, [124] B. Widrow and M. E. Hoff Jr., Adaptive switching circuits, 1960 IRE
1993, pp. 21412144. Western Electric Show Conv. Rec., pt. 4, pp. 96104, Aug. 1960.
[100] A. P. A. Silva, P. C. Nascimento, G. L. Torres, and L. E. B. Silva, An [125] N. Wiener, Cybernetics. Cambridge, MA: MIT Press, 1961.
alternative approach for adaptive real-time control using a nonpara- [126] M. J. Willis, G. A. Montague, D. C. Massimo, A. J. Morris, and M.
metric neural network, in Conf. Rec. IEEE-IAS Annu. Meeting, 1995, T. Tham, Artificial neural networks and their application in process
pp. 17881794. engineering, in IEE Colloq. Neural Networks for Systems: Principles
[101] L. E. B. Silva, B. K. Bose, and J. O. P. Pinto, Recurrent-neural-net- and Applications, 1991, pp. 7174.
work-based implementation of a programmable cascaded low-pass filter [127] M. Wishart and R. G. Harley, Identification and control of induction
used in stator flux synthesis of vector-controlled induction motor drive, machines using artificial neural networks, IEEE Trans. Ind. Applicat.,
IEEE Trans. Ind. Electron., vol. 46, pp. 662665, June 1999. vol. 31, pp. 612619, May/June 1995.
[102] T. Sorsa, H. N. Koivo, and H. Koivisto, Neural networks in process [128] J. M. Zurada, Introduction to Artificial Neural Networks. Boston, MA:
fault diagnosis, IEEE Trans. Syst., Man. Cybern., vol. 21, pp. 815825, PWSKent, 1995.
July/Aug. 1991.
[103] D. F. Specht, Probabilistic neural networks for classification, mapping,
or associative memory, in Proc. IEEE Int. Conf. Neural Networks, July
1988, pp. 525532.
[104] , Probabilistic neural networks, Neural Networks, vol. 3, pp. Magali R. G. Meireles received the B.E. degree
109118, 1990. from the Federal University of Minas Gerais, Belo
[105] A. Srinivasan and C. Batur, Hopfield/ART-1 neural network-based Horizonte, Brazil, in 1986, and the M.Sc. degree
fault detection and isolation, IEEE Trans. Neural Networks, vol. 5, pp. from the Federal Center for Technological Education,
890899, Nov. 1994. Belo Horizonte, Brazil, in 1998, both in electrical
[106] W. E. Staib and R. B. Staib, The intelligence arc furnace controller: A engineering.
neural network electrode position optimization system for the electric She is an Associate Professor in the Mathematics
arc furnace, presented at the IEEE Int. Joint Conf. Neural Networks, and Statistics Department, Pontific Catholic Uni-
New York, NY, 1992. versity of Minas Gerais, Belo Horizonte, Brazil.
[107] R. Steim, Preprocessing data for neural networks, AI Expert, pp. Her research interests include applied artificial
3237, Mar. 1993. intelligence and engineering education. In 2001, she
[108] K. Steinbuch and U. A. W. Piske, Learning matrices and their applica- was a Research Assistant in the Division of Engineering, Colorado School of
tions, IEEE Trans. Electron. Comput., vol. EC-12, pp. 846862, Dec. Mines, Golden, where she conducted research in the Mechatronics Laboratory.
1963.
[109] F. Sun, Z. Sun, and PY. Woo, Neural network-based adaptive controller
design of robotic manipulators with an observer, IEEE Trans. Neural
Networks, vol. 12, pp. 5467, Jan. 2001. Paulo E. M. Almeida (S00) received the B.E. and
[110] M. K. Sundareshan and C. Askew, Neural network-assisted variable M.Sc. degrees from the Federal University of Minas
structure control scheme for control of a flexible manipulator arm, Au- Gerais, Belo Horizonte, Brazil, in 1992 and 1996,
tomatica, vol. 33, no. 9, pp. 16991710, 1997. respectively, both in electrical engineering, and the
[111] R. S. Sutton, Generalization in reinforcement learning: Successful ex- Dr.E. degree from So Paulo University, So Paulo,
amples using sparse coarse coding, in Advances in Neural Informa- Brazil.
tion Processing Systems. Cambridge, MA: MIT Press, 1996, vol. 8, He is an Assistant Professor at the Federal Center
pp. 10381044. for Technological Education of Minas Gerais, Belo
[112] H. H. Szu, Automatic fault recognition by image correlation neural net- Horizonte, Brazil. His research interests are applied
work techniques, IEEE Trans. Ind. Electron., vol. 40, pp. 197208, Apr. artificial intelligence, intelligent control systems, and
1993. industrial automation. In 20002001, he was a Vis-
[113] J. Teeter and M. Y. Chow, Application of functional link neural net- iting Scholar in the Division of Engineering, Colorado School of Mines, Golden,
work to HVAC thermal dynamic system identification, IEEE Trans. where he conducted research in the Mechatronics Laboratory.
Ind. Electron., vol. 45, pp. 170176, Feb. 1998. Dr. Almeida is a member of the Brazilian Automatic Control Society. He re-
[114] L. Tsoukalas and J. Reyes-Jimenez, Hybrid expert system-neural net- ceived a Student Award and a Best Presentation Award from the IEEE Industrial
work methodology for nuclear plant monitoring and diagnostics, in Electronics Society at the 2001 IEEE IECON, held in Denver, CO.
Proc. SPIE Applications of Artificial Intelligence VIII, vol. 1293, Apr.
1990, pp. 10241030.
[115] R. E. Uhrig, Application of artificial neural networks in industrial
technology, in Proc. IEEE Int. Conf. Industrial Technology, 1994, pp.
7377. Marcelo Godoy Simes (S89M95SM98)
[116] A. T. Vemuri and M. M. Polycarpou, Neural-network-based robust fault received the B.S. and M.Sc. degrees in electrical
diagnosis in robotic systems, IEEE Trans. Neural Networks, vol. 8, pp. engineering from the University of So Paulo, So
14101420, Nov. 1997. Paulo, Brazil, in 1985 and 1990, respectively, the
[117] G. K. Venayagamoorthy and R. G. Harley, Experimental studies with a Ph.D. degree in electrical engineering from the
continually online-trained artificial neural network controller for a turbo University of Tennessee, Knoxville, in 1995, and
generator, in Proc. Int. Joint Conf. Neural Networks, vol. 3, Wash- the Livre-Docencia (D.Sc.) degree in mechanical
ington, DC, July 1999, pp. 21582163. engineering from the University of So Paulo, in
[118] B. W. Wah and G. J. Li, A survey on the design of multiprocessing 1998.
systems for artificial intelligence applications, IEEE Trans. Syst., Man, He is currently an Associate Professor at the Col-
Cybern., vol. 19, pp. 667692, July/Aug. 1989. orado School of Mines, Golden, where he is working
[119] S. Weerasooriya and M. A. El-Sharkawi, Identification and control of to establish several research and education activities. His interests are in the
a DC motor using back-propagation neural networks, IEEE Trans. En- research and development of intelligent applications, fuzzy logic and neural
ergy Conversion, vol. 6, pp. 663669, Dec. 1991. networks applications to industrial systems, power electronics, drives, machine
[120] P. J. Werbos, Beyond regression: New tools for prediction and anal- control, and distributed generation systems.
ysis in the behavioral sciences, Ph.D. dissertation, Harvard Univ., Cam- Dr. Simes is a recipient of a National Science Foundation (NSF)Faculty
bridge, MA, 1974. Early Career Development (CAREER) Award, which is the NSFs most presti-
[121] , Maximizing long-term gas industry profits in two minutes in gious award for new faculty members, recognizing activities of teacher/scholars
lotus using neural network methods, IEEE Trans. Syst., Man, Cybern., who are considered most likely to become the academic leaders of the 21st
vol. 19, pp. 315333, Mar./Apr. 1989. century.