DocumentToPDF trial version, to remove this mark, please register this software. ABSTRACT:
pattern matching, predict the outcome from a given set of inputs. The neural
This Report is an introduction to
network trains using a pattern file. In
Artificial Neural networks. It also deals
training, it converges on a proper set of
with an interesting application of neural
weights, or coefficients that lead from
network called “Electronic Nose”.
input
After
training
the
expression that is a function of inputs
being developed as systems for the
and weight coefficients, to obtain the
automated detection and classification of
output.
odors, vapors, and gases. An electronic of
output.
network, simply computes an arithmetic
Electronic/artificial noses are
nose is generally composed
/
a
1.1 A Neural Network
chemical sensing system (e.g., sensor array or spectrometer) and a pattern
An Artificial Neural Network
recognition system (e.g., artificial neural
(ANN) is an information processing
network). We are developing Electronic
paradigm that is inspired by the way
noses for the automated identification of
biological nervous systems, such as the
volatile chemicals for environmental and
brain, process information. The key
medical applications. In this paper, we
element of this paradigm is the novel
briefly describe an electronic nose, show
structure of the information processing
some results from a prototype electronic
system. It is composed of a large number
nose,
of
and
discuss
applications
of
highly
interconnected
processing
electronic noses in the environmental,
elements (neurones) working in unison
medical, and food industries.
to solve specific problems. ANNs, like people, learn by example. An ANN is
1.
Introduction
to
neural
configured for a specific application,
networks
such as pattern recognition or data classification,
DEFINITION:
through
a
learning
process. Learning in biological systems involves adjustments to the synaptic
Neural Networks are form of Artificial
Intelligence
that,
connections that
through
exist
between the
neurones. This is true of ANNs as well.
1
DocumentToPDF trial version, to remove this mark, please register this software. The first artificial neuron was produced
3. Real
Time
Operation:
ANN
in 1943 by the neurophysiologist Warren
computations may be carried out
McCulloch and the logician Walter Pits.
in parallel, and special hardware
But the technology available at that time
devices are being designed and
did not allow them to do too much.
manufactured
which
take
advantage of this capability. 1.2 Why to use neural networks? Neural
networks,
with
4. Fault Tolerance via Redundant Information
their
Coding:
Partial
destruction of a network leads to
remarkable ability to derive meaning
the corresponding degradation of
from complicated or imprecise data, can
performance.
be used to extract patterns and detect
network
trends that are too complex to be noticed
However,
capabilities
some
may
be
retained even with major network
by either humans or other computer
damage.
techniques. A trained neural network can be thought of as an "expert" in the
1.3
category of information it has been given
conventional computers
Neural
networks
versus
to analyze. This expert can then be used to
provide
projections
given
new
Neural networks take a different
situations of interest and answer "what
approach to problem solving than that of
if"questions.
conventional computers. Conventional
Other advantages include:
computers use an algorithmic approach i.e. the computer follows a set of
1. Adaptive learning: An ability to
instructions in order to solve a problem.
learn how to do tasks based on
Unless the specific steps that
the data given for training or
the
computer needs to follow are known the
initial experience.
computer cannot solve the problem. That
2. Self-Organization: An ANN can
restricts the problem solving capability
create its own organization or
of conventional computers to problems
representation of the information
that we already understand and know
it receives during learning time.
how to solve. But computers would be so much more useful if they could do
2
DocumentToPDF trial version, to remove this mark, please register this software. things that we don't exactly know how to
Neural
do.
networks
and
conventional algorithmic computers are not in competition but complement each Neural
networks
process
other. There are tasks are more suited to
information in a similar way the human
an algorithmic approach like arithmetic
brain does. The network is composed of
operations and tasks that are more suited
a large number of highly interconnected
to neural networks. Even more, a large
processing elements (neurons) working
number of tasks, require systems that use
in parallel to solve a specific problem.
a combination of the two approaches
Neural networks learn by example. They
(normally a conventional computer is
cannot be programmed to perform a
used to supervise the neural network) in
specific task. The examples must be
order to perform at maximum efficiency.
selected carefully otherwise useful time is wasted or even worse the network
Neural networks do not perform
might be functioning incorrectly. The
miracles. But if used sensibly they can
disadvantage is that because the network
produce some amazing results
finds out how to solve the problem by itself, its operation can be unpredictable. On the other hand, conventional computers use a cognitive approach to problem solving; the way the problem is to solved must be known and stated in small unambiguous instructions. These instructions are then converted to a high level language program and then into machine code that the computer can FIG: 1 NEURON
understand. These machines are totally predictable; if anything goes wrong is due to a software or hardware fault.
3
DocumentToPDF trial version, to remove this mark, please register this software. 1.4 From Human Neurons to Artificial
to fire (or not), for particular input
Neurons
patterns.
We
conduct
neural
In the using mode, when a
networks by first trying to deduce the
taught input pattern is detected at the
essential features of neurons and their
input, its associated output becomes the
interconnections.
typically
current output. If the input pattern does
program a computer to simulate these
not belong in the taught list of input
features.
patterns, the firing rule is used to
We
However
these
then
because
our
knowledge of neurons is incomplete and
determine whether to fire or not.
our computing power is limited, our models
are
idealizations
necessarily of
real
gross
networks
of
neurons.
FIG: 3 A simple neuron Laboratory FIG: 2 Neuron model
multiprogram
(PNNL).
PNNL
national
is
a
laboratory
operated by Battelle Memorial Institute
1.5 A simple neuron
for the U.S. Department of Energy under Contract DE-AC06-76RLO 1830.
An artificial neuron is a device with many inputs and one output. The neuron has two modes of operation; the training mode and the using mode. In the training mode, the neuron can be trained
4
DocumentToPDF trial version, to remove this mark, please register this software. 2. ELECTRONIC NOSES AND
chemical
THEIR APPLICATIONS
automated fashion difficult.
data
in
an
identification is to build an array of
The two main components of an
sensors, where each sensor in the array is
electronic nose are the sensing system
designed to respond to a specific
and the automated pattern recognition
chemical.
system. The sensing system can be an array
of
several
different
each
element
measures
a
single
sensing
chemical sensors.
(e.g.,
Artificial
networks
analyze complex data and to recognize
can be a combination.
patterns, are showing promising results
Each chemical vapor presented to
in chemical vapor recognition. When an
the sensor array produces a signature or
ANN is combined with a sensor array,
pattern characteristic of the vapor. By
the number of detectable chemicals is
presenting many different chemicals to
generally greater than the number of
the sensor array, a database of signatures
sensors [1]. Also, less selective sensors
is built up. This database of labeled
which are generally less expensive can
signatures is used to train the pattern
be used with this approach. Once the
recognition system. The goal of this
ANN is trained for chemical vapor
training process is to configure the
recognition,
recognition system to produce unique
identification quantity
operation
consists
of
propagating the sensor data through the
classifications of each chemical so that
The
neural
(ANNs), which have been used to
measurements for each chemical, or it
beimplemented.
the
and difficult to build highly selective
spectrometer) that produces an array of
automated
approach,
being monitored. It is both expensive
different
device
this
least as great as the number of chemicals
property of the sensed chemical, or it can a
With
number of unique sensors must be at
sensing
elements (e.g.,chemical sensors), where
an
of
One approach to chemical vapor
.
be
analysis
network. Since this is simply a series of
can
vector- matrix multiplications, unknown
and
chemicals can be rapidly identified in the
complexity of the data collected by
field. Electronic noses that incorporate
sensors array can make conventional
5
DocumentToPDF trial version, to remove this mark, please register this software. ANNs
have
standard
multilayer
various applications.
network
trained
Some of these applications will be
backpropagation algorithm and the fuzzy
discussed later in the paper. Many ANN
ARTmap
configurations and training algorithms
operation a chemical vapor is blown
have been used to build electronic noses
across the array, the sensor signals are
including backpropagation-trained, feed-
digitized and fed into the computer, and
forward
the ANN (implemented in software) then
demonstrated
in
networks; fuzzy ARTmaps;
Kohonen’s (SOMs);
been
self-organizing learning
vector
maps
with
algorithm
identifies
the
feed-forward
[2].
chemical.
the
During
This
quantizers
identification time is limited only by the
(LVQs); Hamming networks; Boltzmann
response time of the chemical sensors,
machines; and Hopfield networks.
which is on the order of seconds. This
Figure 1 illustrates the basic schematic
prototype nose has been used to identify
of an electronic nose.
common household chemicals by their odor [3].
FIG:4 Schematic diagram of EN 2.1
PROTOTYPE
FIG: 5 Photograph of the prototype
ELECTRONIC
electronic nose
NOSE One of our prototype electronic
Figure 3 illustrates the structure of the
noses, shown in Figure 2, is composed
ANN. The nine tin-oxide sensors are
of an array of nine tinoxide vapor
commercially
sensors, a humidity sensor, and a
gas sensors obtained from Figaro Co.
temperature sensor coupled with an
Ltd. (Sensor 1, TGS 109; Sensors 2 and
ANN.
3, TGS 822; Sensor 4, TGS 813; Sensor
available
Taguchi-type
were
5, TGS 821; Sensor 6, TGS 824; Sensor
constructed for this prototype: the
7, TGS 825; Sensor 8, TGS 842; and
Two
types
of
ANNs
6
DocumentToPDF trial version, to remove this mark, please register this software. Sensor 9, TGS 880). Exposure of a tin-
are presented to the system. By training
oxide sensor to a vapor produces a large
on samples of various chemicals, the
change in its electrical resistance.
ANN learns to recognize the different
The humidity sensor (Sensor 10: NH-02)
chemicals. This prototype nose has been
and the temperature sensor (Sensors 11:
tested on a variety of household and
5KD-5)
office
are
used
to
monitor
the
supply
chemicals
ammonia,
including
conditions of the experiment and are also
acetone,
ethanol,
glass
fed into the ANN.
cleaner, contact cement, correction fluid, iso-propanol, lighter fluid, methanol, rubber cement and vinegar. For the results shown in the paper, five of these chemicals were used: acetone, ammonia, isopropanol, lighter fluid, and vinegar. Another category, “none” was used to denote the absence of all chemicals except those normally found in the air which resulted in six output categories from the ANN. Table 1 lists the training Parameters for one backpropagation and
FIG:6
Structure
of
one fuzzy ARTmap network.
the
Backpropagation
backpropagation ANN used in the prototype
to
Architecture: 11-11-6 feedforward
identifyhousehold
Activation: Logistic Sigmoidal
chemicals
Learning Rate: 0.10
Although each sensor is designed
Momentum: 0.90
for a specific chemical, each responds to a
wide
variety
of
No. of Epochs: 1369
chemicals.
Fuzzy ARTMap
Collectively, these sensors respond with
Training Vigilance: 0.98
unique signatures (patterns) to different
Testing Vigilance: 0.80
chemicals. During
No. of Epochs: 3 the
training
process,
ANN training parameters
various chemicals with known mixtures 7
DocumentToPDF trial version, to remove this mark, please register this software. Both
networks
were
trained
using
Fluid Vinegar
randomly selected sample sensor inputs.
66
21
The ANNs used here were not trained to
68
26
1
2
the analytes. This allowed the ANN to
Table
1:
This paper was presented at the IEEE
backpropagation
Northcon/Technical
ARTmap (FA)
quantify the concentration level of the
Amm & Vinegar Isopr & Vinegar
identified analytes, but were trained with
92.6 81.0 95.2 92.3 76.9 00.0 00.0
samples with different concentrations of
Applications
ANN
performance (BP)
and
for fuzzy
Conference (TAC’95) in Portland, OR,
Figures 4 and 5 illustrate the responses
USA on 12 October 1995. generalize
of
well on the test data set. Performance
classification for a variety of test
levels
chemicals presented to the ANNs. The
of the
two
networks were
the
sensors
and
the
ANN
basically equivalent ranging from 89.7%
ANN was able to correctly classify the
to 98.2% correct identification on the
test samples with only small residual
test set depending on the random
errors. While the ANN used here was
selection of training patterns. Table 2
not trained to quantify the concentration
summarizes
level of the identified analytes, it was
one
set
of
network
trained with samples with different
performances for novel sensor inputs.
concentrations of the analytes. This Num Train
Num Test
Input Substance
67
28
None
75
22
Acetone
64
14
Ammonia
93
28
Isopropanol
5
3
106
25
Ammonia &Isopr Lighter Fluid Amm & Lig
74
27
% Correct BP FA 96.4 96.4 100 100 100 100 92.9 100 00.0 66.7 100 96.0 100
allowed the ANN to generalize well on the test data set. From the responses of the sensors to the analytes, one can easily see that the individual sensors in the array are not selective (Figure 4). In addition, when a mixture of two or more chemicals is presented to the sensor array, the resultant
pattern (sensor
values) may be even harder to analyze (see Figure 5: c, d, and e). Thus, analyzing
8
the
sensor
responses
DocumentToPDF trial version, to remove this mark, please register this software. separately may not be adequate to yield
2.3
ELECTRONIC
the classification accuracy achieved by
MEDICINE
analyzing the data in parallel.
NOSES
FOR
Because the sense of smell is an important sense to the physician, an
2.2 ELECTRONIC NOSES FOR
electronic nose has applicability as a
ENVIRONMENTAL MONITORING
diagnostic tool. An electronic nose can
Enormous amounts of hazardous waste
examine odors from the body (e.g.,
(nuclear, chemical, and mixed wastes)
breath, wounds, body fluids, etc.) and
were generated by more than 40 years of
identify possible problems. Odors in the
weapons’
breath
production
in
the
U.S.
can
be
indicative
of
Department of Energy’s weapons’
gastrointestinal problems, sinus
complex.
Northwest
problems, infections, diabetes, and liver
National Laboratory is exploring the
problems. Infected wounds and tissues
technologies
perform
emit distinctive odors that can be
environmental restoration and waste
detected by an electronic nose. Odors
management in a cost effective manner.
coming from body fluids can indicate
This effort includes the development of
liver and bladder problems. Currently,
portable, inexpensive systems capable of
an electronic nose for examining wound
real-time identification of contaminants
infections is being tested at South
in the field. Electronic noses fit this
Manchester University Hospital.
The
Pacific
required
to
category.
A more futuristic application of
Environmental
applications
of
electronic
noses
has
been
recently
electronic noses include analysis of fuel
proposed for telesurgery. While the
mixtures [4], detection of oil leaks [5],
inclusion of visual, aural, and tactile
testing ground water for odors, and
senses
identification of household odors [3].
widespread, the sense of smell has been
Potential
include
largely ignored. An electronic nose will
identification of toxic wastes, air quality
potentially be a key component in an
monitoring, and monitoring factory
olfactory input to telepresent virtual
emissions.
reality systems including telesurgery.
applications
9
into
telepresent
systems
is
DocumentToPDF trial version, to remove this mark, please register this software. The
electronic
nose
3. CONCLUSION
would
identify odors in the remote surgical environment.
These identified
Thus
an
“Artificial
Neural
odors
Network” is developed to make the
would then be electronically transmitted
computer think like a human brain. And
to another
an electronic nose is a device intended to
site where an odor generation system
detect odors or flavors. Over the last
would recreate them.
decade, “electronic sensing” or “e-
2.4 ELECTRONIC NOSES FOR THE
sensing” technologies have undergone
FOOD INDUSTRY
important developments from a technical
Currently, the biggest market for
and commercial point of view. The
electronic noses is the food industry.
expression “electronic sensing” refers to
Applications of electronic noses in the
the capability of reproducing human
food industry include quality assessment
senses using sensor arrays and pattern
in food production , inspection of food
recognition systems. For the last 15
quality by odor, control of food cooking
years as of 2007, research has been
processes, inspection of fish, monitoring
conducted
the
checking
commonly referred to as electronic
rancidity of mayonnaise, verifying if
noses, that could detect and recognize
orange juice is natural, monitoring food
odors and flavors.
fermentation
process,
and beverage odors, grading whiskey, inspection
of
beverage
to
develop
technologies,
These devices have undergone
containers,
much development and are now used to
checking plastic wrap for containment of
fulfill industrial needs.
onion odor, and automated flavor control to name a few. In some instances electronic noses can be used to augment or replace panels of human experts. In other cases, electronic noses can be used to reduce the amount of analytical chemistry that is performed in food production especially when qualitative results will do.
10
DocumentToPDF trial version, to remove this mark, please register this software. REFERENCES [1] B.S. Hoffheins, Using Sensor Arrays and Pattern Recognition to Identify Organic Compounds. MS-Thesis,
The
University
of
Tennessee, Knoxville, TN, 1989. [2] G.A. Carpenter, S. Grossberg, N. Markuzon, J.H. Reynolds, and D.B. Rosen, “Fuzzy ARTMAP: A Neural
Network
Incremental
Architecture
Supervised
Learning
for of
Analog Multidimensional Maps,” IEEE Transactions on Neural Networks, vol. 3, 698 -713. [3] P.E. Keller, R.T. Kouzes, and L.J. Kangas, “Three Neural Network Based Sensor
Systems
Monitoring,”
for
IEEE
Environmental Electro
94
Conference Proceedings, Boston, MA, 1994, pp. 377-382. [4] R.J. Lauf and B.S. Hoffheins, “Analysis of Liquid Fuels Using a Gas Sensor Array,” Fuel , vol. 70, pp. 935-940, 1991. [5] H.V. Shurmur, “The fifth sense: on the scent of the electronic nose,” IEE Review, pp. 95-58,
11