1. Neural networks are being used in manufacturing to optimize steel production processes and control systems. Historical plant operations data that would otherwise sit unused can be utilized by neural networks as online knowledge for artificial intelligence-based process control.
2. Specifically, a steel plant in Western India developed two neural network models to optimize the feed mix ratio for sponge iron production. The models predict raw material quantities, production consumables needed, and sponge iron output based on over 40 input parameters like material costs and qualities.
3. The neural network models were able to accurately match historical plant data, enabling improved process control, economic analysis, and "what-if" scenario planning to optimize production costs and output.
IRJET- Effect of ICT Application in Manufacturing Industry
NeuralNetworks AI DCS
1. TIIANUFACTURING
Produetion lllanagcmcnt
Ncural netrilorlrs automa tc @-mix
optimisation in stcel production
Historical data from plant
operations can now serye as
on-line knowledge for
artificial intelligence based
control. Manufacturers are
reaping rich rewards through
the application of neural
networks to automation
P R Pnqsno B VrneNorn StNou
I rtificial intelligence is making
l- its presence felt in a variety
of manufacturing applications.
Automation systems and industrial
software manufacturers have begun
"pplyrg
a host of new technologies
such as fiv,zylogic and neural netrvorks
to get a better grip on processes.
Machines no longer suffice it seems.
Machines with brains, albeit anificially
generated ones, do. Increasingly,
artificial intelligence techniques are
being employed to resolve many a
quirk in the production process. End-
product quality, raw mate rial
inventories, their cost and production
output all have a significant bearing
on business. Each of these variables
need to be tightly monitored and
controlled. The compulsions to go
beyond traditional control to
optimisation of the process is a given
in these circumstances.
P R Prasad is Project Manager at Bendy
Nevada ( India) Pvt. I td., Mumbai. He
holds a Ph.D in Chemical Engineering and
has over ten years experience in process
automation, advanced control and anificial
intelligence applications.
Virender Singh is a Senior Engineer with
KLG Sptel Ltd., and holds a Bachelor's
degree in Computer Science.
56
Whcn aduanced eontrol
becomcs a nust
Tladitional PID control has severe
limitations when it comes to
multivariable, non-linear and random
processes. Flere, advanced control
techniques need to be used. Very
often, the challenge of controlling
difficult loops is more a matter of
uncertain data rather than uncertain
process behaviour. In such situations,
artifi cial intelligence techniques using
neural networks help to infer the
current value of process variables from
a collection of related variables such
as the historical data of the process.
Neural net'work technology has been
successfully used in a variety of
process control and optimisation
applications for the last few years.
Artificial neural nerworks, or neural
networks for short are inspired by
biological neural networks found in the
brain of mammals. The mathematical
formulation of neural networks
enables them to "learn" and "organise"
themselves, i.e. adjust various
computational parameters so that the
characteristics of the data are captured
within the neural network formulation.
Neural networks thus can adapt to
changes in data and learn to generalise
the characteristics of input data. For
example, a neural nemrork can learn
the mapping between an input and
output set of data, such that it predicts
the appropriate ouq>ut when presented
with a new input data. This property
can be effectively used in a variety of
tilll r December 1998
2. practical Pattern-recognition and
classifi cation applications
Puttlng plant hlstort to usc
In several plants or production
processes, there are a large number of
variables that affect
the production
(output) or product
quality.Therefore, it
is generally diffrcult
to choose the right
combination of all
these variables to
ensure optimality -
that is, maximum
production or best
product quality.
Sometimes, the
effects of the
numerous varia-
bles on optimised
operations are not
fully understood.
Usually a large
arnount of historical
data pertaining to
the operation of the
plant and the values
of the numerous
variables that were
used in the past is
available in plant
logbooks or in the Distributed Control
System. Such historical data can be
readily classified as "near-optimal
operation", "good oPeration" and
"bad operation". Neural networks can
be applied on such historical data to
"learn" the pattern of the variables that
provide the above categories of
operation. Once the network has learnt
these patterns, it is possible to use it in
a predictive mode to identify which
combination ofvariables to use in order
to achieve a desired production or
product quality. A case study of how a
neural network based application was
developed for a steel manufacturer is
discussed in the next section.
Fccd nlr optlnlsatlon
The problem of feed mix optimisation
in a steel plant is one of choosing the
"optimal" feed mix ratio of different
iron ores for sponge iron production.
This choice of optimal feed mix ratio
is made by taking into consideration
the following factors: inventory of raw
materials, cost of raw materials, cost
of production consumables.
The first thing to do in developing a
neural network based solution is to
identify all the factors i.e, input
parameters that affect the desired
output values. For example, in the case
of steel production at a major plant site
in western India, the following outPuts
were to be predicted by the neural
network based system using Gensym's
G2 and NeurOn-Line software :
I Quantities of input iron ores
I Quantities of production
consumables needed to Process the
above quantities of iron ore
I Quantity of sponge iron product
that can be produced with the feed mix.
After several discussions with plant
personnel, the input Parameters that
were identified as affecting the outputs
were the cost of various iron ores)
their inventory qualiry parameters of
iron ores and the desired quality of
sponge iron product.
A carefirl analysis ofthe inputs and
outputs identified above revealed that
there are over fotty inputs that affect
the desired output values. This is
because there are five different tyPes
of iron ores. For each of these iron
ores there are six different quality
parameters that need to be taken into
account, thus amounting to thirty
inputs. The rawmaterial cost (5), raw
material inventory (5) and product
quality parameters (3) work out to
another thirteen inputs. There are
nine output values that are to be
predicted by the neural network
model - five raw material quantities,
three production consurnable values
and one sponge iron product quantity.
Instead ofdevelopitg
"
massive neural
network that takes into account 43
inputs and 9 outputs, the problem was
decomposed into two sub-problems,
each ofwhich had a lesser number of
inputs and outputs as shown
rn Figure I.
Dete collcctlon C trelnlng
Once the above design was
complete and approved by the plant
personnel, then historical data for the
above inputs and outputs were
collected from the plant DCS and log
books and manually filtered. By
filtering, we mean removing input
and output data sets for those daYs
when some condition was not normal
Production
ModelRaw Material
Quantity Predictor
ilfll r December 1998
57
3. ilAilUFACTURING
Production frlanagcncnt
- for example, part of the plant may
have been shutdown or not fully
operational, product quality on
specific days may have been very poor
due to factors not necessarily under
the control ofthe plantpersonnel, some
recorded data may be incorrect due
to known instrumentation errors) etc.
Only "good data" after manual
filtering from the previous six months
of operation was collected. This data
was then used for "training" -
adjusting the parameters - of the rwo
neural networks. While the details of
the training algorithm are out of the
scope of such an introductory article,
it must be noted that the algorithm
itself is fairly easy to undersrand and
easily implementable.
Rcsults cf thc ncrtal
nctuork nodclllng
Using the data collected the two
neural network based models were
trained and excellenr modelling results
were obtained. The modelling results
for the production model (as described
in Ftgure 1) are illustrated in ttgure
2. The solid line in Fbuo 2 repre-
sents the actual historical data as col-
lected from the plant. The dots on the
plots represent the predictions made
by the neural network given the vari-
ous inputs. The close correspondence
between the historical data and the val-
ues predicted by the neural network
gives ample proof that the model de-
veloped is reasonably accurate. A host
of other benefits accrue to the produc-
tion departrnent as a result of the neu-
ral network implementation. These
include :
Non-Linear Models: The neu-
ral network models built during this
application have captured the operat-
ing practices of the steel plant. Given
a set of raw material inventory qual-
ity, costs and desired product quality
parameters, the two neural network
models can predict the raw material
quantities required to be used and the
arnount of production output and pro-
duction consumables required. The
production personnel can figure out
58
how much quantiry of raw materials
is to be used for achieving a desired
production target. Further, the pro-
duction supervisors have an estimate
production model to predict the im-
pact on production and
production consrunables, and thereby
derive the impact on cost of sponge
iron production. Such what-if analy-
ses are precisely
what the higher
management
desired in order
to set oPera-
tional targets
for the pl.""l Uy
maxlmrslng
production and
minimis-
ing cost.
This project
was imple-
mented using
Gensym Corpo-
ration's
NeurOn-Line
software tool.
The user inter-
face that has
Figure 2: Rcsults of ncural network nodelling
0,0
0,D
0,0
0.0
The results of the neural network predictions shows a
strikingly close correspondencewith the training data.
of the required production
consumables.
Economic Analysis: Orlce the raw
material and product quantities are
known as explained above, it is a sim-
ple matter to plug in the actual costs
for the raw materials and product and
do an economic analysis of produc-
tion cost per ton of sponge iron. The
management thus has an easy-to-use
been provided for the end-user is
shown 1n
Figure 3. As is evident from
Figure 3, the use of the neural net-
work technology is completely trans-
parent to the end-user. All that the user
has to do is to choose different
amounts of raw materials using the
sliders shown in Figure 3. The raw
material quality data and other inputs
tool to do an
economic analy-
sis of the pro-
duction process.
What-if
Analysis: Since
the models rep-
resent a reason-
ably accurate be-
haviour of the
plant, it is pos-
sible to use these
models to per-
form what-if
analyses. That
is, various quan-
tities of raw ma-
terial inputs can
be input to the
Neural network technology enables implementation of "what-if "
ana lysis th rough tra nsparent user- i ntefaces.
Figure 3: Uscr intcrfasc prouidcd bl l{curOn-linc
ilm r December 1998
4. iIANUFACIURII{G
to the neural network models are read
from a file that is stored in the
computer. The results of the neural
network pre diction, that is,
the sponge iron Produced, the
consumablesrequired and the cost of
production are all displayed in the
form of dials for easy visualisation.
networks work
follows:
The XOR Logic Gate is implemented in
the form of a neural network. Here, neurons
A and B are the inPut neurons. Neurons la-
belled HI and H2 represent the hidden layer
of neurons. The output layer of the network
Inv(0.5) to indicate that the step function
has a shape that is exacdy tlre inverse of that
shown above. Flence, if the input to this
threshold function is less than 0.5, the out-
put is unity; else, if the input is greater than
0.5, the output is zero.
lrt us assurne that the input pattern Pre-
sented to this network is A = I, B : 0.
Then the internal state of the neuron 'HI' is
computed as follows:
(-l) (l) + (r) (0): -r
Since -I< 0.5 (the threshold limit of the
hidden neuron HI), the outPut of neuron
Hl is zero. The internal state of the hidden
neuron H2 is computed as:
(-I) (I) + (r) (0) : -I.
Since -I< - 0.5 ( the threshold of H2), the
output of neuron H2 is zmo.
Now with both HI and H2 outPuts at zrro)
the internal state of neuron 'O' is given as:
(0.5) (0) + (0.5) (0) : o
Since 0 < 0.5 ( the threshold limit of the
inverse threshold function), the output from
neuron 'O' is unity, exacdy what is expected
from an XOR gate. It must be noted that the
XOR gate cannot be implemented in the
form of a neural network if the hidden lay< r
is not present. In other words, the hidden
layer is a necessary part of the network ar-
chitecture if non-linear functions are to be
mapped.
This example illustrates how neural net-
works store knowledge in the form of
weights of their connections. For compli-
cated patterns, there would be a larger
number of neurons used - both in terms of
the number of layers of neurons and also
the number of neurons in each layer. Thus,
only the form of the network changes de-
pending on the problem in hand.
The ptor tnd cons
The technology of neural networks
is still at an early stage of application'
considering that it was only in 1987
that the first commercial aPPlications
of neural networks were rePorted.
Despite this there has been a raPid and
widispread exploitation of the
technology for solving a variety of
problems. The reasons for this
acceptance over multiple domains are
because of the wonderfial benefits that
the technology ofFers. The benefits of
a technology can be better understood
with respect to a specific application
as is apparent from the above case'
When used as a soft sensor or virnral
sensor some of the advantages are:
I The ability to Predict current
or future values using models.
r Avoid laboratory test delaYs if
used as a virfttal sensor and thereby
enhance Process control.
I Identify key influencing
variables: M*y times the sensitivity
of an output to an inPut variable
changes dramatically depending on the
operating region. An off-line model
can predict the sensitivity at the
current operating conditions and not
just an average sensitivity over all
operating points.
I Provide a non-linear model for
Process control.
I Sensor validation: A unit may
already have an on-line sensor, but the
sensor may be subject to frequent er-
rors (bias, drift, dead sensor) and need
frequent maintenance.
A virtual sensor or inferential
sensor can cross-check and validate
the on-line analyser, Providing a
means of detecting when a sensor
becomes f"olty.
Its internal state is given as the weighted
sum of inputs from the previous layer and its
bias. In the brain, a tremendous number of
neurons are interconnected to form a net-
work which is capable of performing ad-
vanced
An
by con
neuron
belorv illustrates a wpical neural network
with three lavers - input, hidden and out-
Inputs
neural net-
work. The ac-
tual result ex-
pected for dif-
ferent sets of
lnPuts to a
XOR gate is as
put / as is shown below.
How ncural
Output
put layers. When presented with a set of
inpot-output data, the neu'ral network
"learns" the patterns by adjusting the weights
of each neuron. Connections (i.e. weights,
w,) between neurons which lead to the "right
answer" are strengthened (i.e. increased in
magnitude) while those leading to "incor-
rect answers" are decreased by repeated ex-
posure to sample problems (i.e. the data
iets). Thus, neural network have the ability
to learn in a manner closely parallel to hu-
man learning. Knowledge is thus stored by
the strengths of the interconnection be-
tween neurons.
To understand how a neural network
executes and obtains its results, Iet us con-
sider a simple example, i.e. how to imple-
ment a XOR logic gate in the form of a
0
1
't
0
5. Production frlanagcncnt
lnfurcntial scnsors
Also known as Virnral Sensor or Soft
Sensor, an inferential sensor is a software
package that uses a neural network to
compute values of a variable based on
other related data inputs. VItrhile a physical
sensor direcdy measures a value, i virtual
sensor predicts the same value using other
values as inputs to a neural networ[ based
model. For e
. stream exiting
lJ;ot
directlY
lyser is not
avai
imP Procure or
the neural network technology to learn
the pattern between the input data, i.e.
olumn,
strearn.
virtual
I Zero maintenance: Due to the
use ofsoftrvare to do the measurement,
there is hardly any physical
maintenance associated with respect
to the sensor itself.
Similar advantages can be identified
for almost any other application of
neural network technology However,
neural networks have their share of
disadvantages.
Neural network technology depends
heavily on the availability of "large,,
amounts of "good" data for training
the network. By "good" data, we
mean data that is representative of the
entire sample space under
consideration. If data covering some
region of interest were not available,
then the network would not be able
to identify panerns belonging to that
reglon.
Put simply, neural nerworks are fairly
good if used to "interpolate" values
within the range of training set data.
Ffowever, their results outside the
training range are complete ly
unknown. Further, there should be
sufficiendy large number of exemplar
data to determine a solution set of
60
weights for the network. Otherwise,
an under-trained network may not
be able to "generalise" information
from the data presented to it.
lssucs bcforc thc
production nanagcncnt
As has been pointed out earlier, a
neural network application can be de-
veloped only if there is a "large',
amount of "good" data that is repre-
sentative of the domain being mod-
elled. If historical data is unavailable
or if the data does not reflect the cur-
rent operational state of a plant, then
such data will be useless as far as neu-
ral network modelling is concerned.
Choice of software tools
Several vendors have been making
the neural network technology avail-
able for general use. Pavilion Technolo-
gies Inc., USA, has been touting sev-
eral success stories about neural net-
work applications in the process indus-
try using their software package called
Process Insight. Similarly, Gensym
Corporation has a product called
NeurOn-Line which has been put to
use in a wide varietv of industries.
Also, some Distribut"d Cont ol Sys-
tem vendors are making these tech-
nologies available direcdyon the DCS.
For example, Fisher-Rosemount has
a tool called Intelligenr Sensor Toolkit
that enables users to develop inferen-
tial sensing applications directly on
the DCS itself.
Using these software packages, the
details of the network learning proc-
ess and associated weight pattern in
which the knowledge resides are in-
visible, and in reality, unnecessary for
the user to know for using the sys-
tem effectivelv.
It is extremely important to take an
informed decision about the specific
software tool to be used for a particu-
lar project. For example, using
Gensym's NeurOn-Line product, it is
possible to implement an application
that can communicate with the con-
trol system in real-time.
On the other hand, if you have an
existing DCS for which such a soft-
ware tool is just an add-on feature, then
it is better to implement the applica-
tion using such tools. The cost of the
Iflaior applications of neural nctworlrs
sensors has been in the area of
Continuous Emission Monitoring
Systems. Neural network based virtuJ
ttre US
cy as an
sm for
variety.of industries are for measuring the
following properties:
Octane Number, Melt Index of
p_ornts rnto a glven set of output points. The Intelligent Sensor Toolkit on a Fisher PROVOX
When a neural network is tiainid and control system allows darelopment of virtual sensor:
llll r n^^^--L---
6. in process nanufasturing
the given inputs to the desired outputs.
Thus neural networks are tools for non-
Iinear regression.
Once a non-linear model of a process
is obtained, then the model may be used
to predict the behaviour of the process in
Iuture.
A large number of applications of
neural networks have been reported in
non-linear modelling. Some examples
include: Reactor Models, Autoclave
Models, Furnace Models, Dispersion
Models, Distillation Column Models,
Fluidised Catalytic Cracking lJnit,
Fermentation Process End Time
Prediction and Load Forecasting.
Aduanccd control
Non-linear models, as described in the
previous sub-section, can form the basis
for model based control of highlv non-
linear processes, which generallv cannot
be tackled by classical model based
control techniques. A classic example of
a non-linear process is the neutralisation
software tool, the mode of use of the
application (whether on-line or off-
line), flexibility of application develop-
ment, the user interface ar-ailable are
all factors fhat one should s-eigh c.are-
firlly before taking a decision on the
software tool.
Skilled Personnel
As mentioned above, the availabilio'
of sofrware tools has obviated the ne-
cessity of having specialists u.ho
have a deep understanding of the
technology.
Nevertheless, it is important that the
application developer has a general un-
derstanding of the technology so that
he can recognise commonly known
problems during training of the
network.
For example, there are several issues
with respect to neural nerwork design
- choice of the number of hidden lay-
ers, the number of nodes in each hid-
den layer, the learning parameter, the
use of an appropriate learning algo-
rithm, recognising when a network has
begun "memorising" input-output
sets of data.
llttl r December 1998
IIIAI{UFACru$NG
Produstion frlanagcnent
Predicting pH dynamics with the NeutOn-Line software from
Gensym Corpn.
process where the need is to control pH at
7 by manipulating either addition of an
acid or alkali.
The response of pH to the addition of
acid or base is highly non-linear. Several
Retraining systems
Dependrng on the nature of the ap-
plication, it may be necessary; in some
cases, to have the network re-trained
s'hen new sets of data are obtained.
This is because the operational data drat
s-as used during the initial training
phase mav no longer be valid or may
need to be upgraded to include new
operational regimes encountered more
recendr: As u'as pointed out in the pre-
vious section, neural nem'orks are good
onlv for interpolation within the op-
erational regime covered by data in the
training set. If during any stage of the
use of the technology it is found that
the neural network performance has de-
teriorated, itgenerally points to the fact
that the network needs re-training us-
ing a new set of data.
Implementation and payback
The first neural network application
being developed in an organisation can
take as much as 3-4 months before a
working protot)?e is ready For on-line
applications, the development time may
be a litde longer depending on the types
of interfaces with control systems that
success stories have
been reported in the
literature where the
neutralisation process
can be modelled using
neural neworks. The
prediction from the
neural network based
model is used as a basis
for selecting the
control move to
implement. Advanced
Intelligent Control
using neural networks
includes
r Supervised
Control to learn
mapping between
sensor inputs and
desired actions
r Inverse C-ontrolwherethe network
learns the inverse dvnamics of svstem.
r Neural Adaptive Control for
predicting future process outputs and
control based on predicted error.
are desired. The pay back time for some
of these applications - such as the vir-
tual sensors application - can be easily
computed based on reduced mainte-
nance, faster response time, better prod-
uct qualiry etc. Flowever, in other ap-
plications, such as the case study de-
scribed earlier in this article, where the
network models are used as decision-
support systems, the pavback time may
be more difficult to gaLrge or compute.
Conclusion
The neural network technology has
been around long enough but the pri-
mary impediment to their widespread
use has been the steep learning crrrye
associated with developing applications
by using this. However, with several
vendors making this technology avail-
able for use in an extremelyuser-friendly
manner) applylng it has become as sim-
ple as using a spreadsheet on a PC. The
ability of these networks to learn and
adapt, self-organise and approxirnate
frrnction mappings makes them attrac-
tive for solving fairly complex problems
of automating production.
EU
61