FIELD OF THE INVENTION
The present invention relates to an improved casting advance fault forecasting
system to prevent breakouts in continuous casting process. More particularly, the
present invention relates to a technique for online automated adaptation of
casting advance fault forecasting system to prevent breakouts in a thin steel-slab
caster under drifting conditions.
BACKGROUND OF THE INVENTION
The steel making chain of processes initiates with mineral extraction and
refinement, passing through coke and sinter making, making of liquid iron in a
blast furnace, followed by its transformation to liquid steel through chemical
processing. The next major step is the continuous transformation of liquid steel
to solid, followed by thermo-mechanical processing of the produced discrete solid
slabs (or blooms/billets) into thin sheets of different sizes. Among all these steps,
the continuous transformation of liquid steel to solid is the most critical step both
from the perspective of product quality as well as process stability including
throughput. Any disruption, delay or slackness in the continuous casting stage
immediately exhibit negative impact both on the upstream liquid and
downstream solid processing. The continuous transformation of liquid steel is
called continuous casting on which this invention is focused. More information on
continuous casting may be found in [1,2].
One of the typical and most undesirable disruptions of the continuous casting
process is due to occurrence of "breakouts" in the mold. These occur when the
thin solid shell that develops all around the perimeter of the continuously cast
strand thickness with the progression of cooling, and breaks for some reason

resulting in emptying out of the liquid steel onto the downstream machinery and
equipment effectively disrupting the casting process for a few hours. There are
many other causes of breakouts for example, primarily stickers, cracks and
improper shell formation. Their relative frequency depends on the specific nature
of the casting process.
Conventional slab casters (also called thick slab casters) come with slab sectional
thickness of around 200-250 mm, and capable of casting at a speed between
1.2-2.2 meters per minute. In modern thin slab casters (figure 1 provides a
schematic of the complete thin slab casting line), the slab sectional thickness and
casting speed vary between 75-95 mm and 4-6 m/min. Breakouts in the former
are predominantly of the "sticker" type where formative thin shell sticks to the
mold wall and the surrounding region is torn apart due to tension. In the latter
type of casters, the breakouts are predominantly caused by cracks and other
shell formative defects like "lumb" where shell is weak at the narrow faces,
bulges and breaks as the ferro-static pressure increases.
Considering the seriousness of disruption in the steel making chain caused by
breakouts, attempts have been made since early days of conventional casting to
develop methods to detect formation of breakout-like conditions and take
consequent action to prevent the breakout. For this purpose mold walls are
embedded with arrays of horizontally layered thermocouples (TC), each TC
reflecting the instantaneous variation in temperature of strand shell closest to its
sensing point. "Stickers" are accompanied by a typical rise and then fall of
temperature as the "tear" passes the point, and then a corresponding rise and
fall in the TC placed immediately below. This typical pattern with occasional
(sometimes severe) variations is detected by a pattern-recognition software that
alarms the formation of a tear which is likely to lead to a breakout within a short

period, the alarm is followed by immediate reduction in casting speed allowing
the fault (tear) to heal and thus prevent the sticker breakout (Indian Patent
Appl.420/KOL/04).
While the (thermocouple) TC based breakout detection system works well in
conventional casters, they do not adequately serve the purpose in thin slab
casters for the following two reasons:
(1) unlike conventional casters, breakouts of sticker type constitute only a minor
proportion of all the types of breakouts. Further, the other types of breakout
more typical to thin casters are not accompanied with neat temperature-time
patterns as seen in the stickers type of break-out.
(2) due to higher casting speeds, the residence time of a fault in the mold, after
its detection by the TC, is only around 10 secs. The speed reduction takes effect
in around 6 secs, leaving very little time for actual healing even after detection.
Thus even if a fault were detected by TC in the mold, due to time constraint
there is no guarantee that the resultant breakout can be prevented.
These two factors make the conventional TC-based breakout detection, a poor
method for thin slab casters, and alternatives need to be found.
The second disadvantage as clarified hereinabove, makes it more imperative that
the fault be detected even before it has formed in the mold, because once it
forms there will be little time left to save the situation. This eliminates the
possibility of using TC for direct one-to-one mappings between temperature-time
patterns and crystallized faults. Furthermore, the detection can only be made by
recognizing some typical emergent pattern prior to fault formation in a few

among the hundreds of online parameters that characterize continuous casting.
This is on the expectation that a fault formation is preceded by, or in other
words the root cause of a fault lies in, a certain specific combinatorial pattern
acquired by some among the hundreds of parameters characterizing casting.
This specific combinatorial pattern has to be recognized in process real time by
the new method. Over and above, it is imperative that the recognition be
accurate in the sense that nearly all genuine faults be captured, while practically
no false alarms are raised. And the alarming has to be in sufficient advance to
allows definitive operator response for mitigation, which implies at least three
minutes in advance of a breakout.
Examples of prior work on thermocouple based breakout detection system may
be seen in IPA 420/KOL/04, while examples of prior work based on multi-
parametric analysis are seen in US patent 6,564,119 and US 2004/0172153. In
the latter statistical techniques are used to evaluate the difference or 'distance'
between a set of running parameters of casting and a corresponding set of ideal
parameters extracted from best conditions. The distance provides real time
monitoring and when it exceeds some threshold an alarm is raised. According to
the prior art, the running conditions of casting are taken as the baseline and
significant deteriorations as predicted by an Artificial Neural Network (ANN)
trained for this purpose, are considered as digressions which may lead to
breakout. While the ANN output can be viewed as a monitoring parameter,
severe digressions are alarmed for immediate responsive action.
The ANN used in the (CAFOS), casting advance fault forcasting system is trained
from data extracted from a certain period (months or even years) of casting. It
works accurately as long as the running process conditions are within the bounds
that contain the input and output variable that have been used to train this ANN.

However, industrial processes drift, and very often the process conditions
extrude these bounds. Under such extrusions the accuracy of ANN prediction
gets affected and the reliability of predictions from CAFOS is severely reduced.
Hence, it becomes essential to develop a mechanism which can address process
drift, and more specifically, adapt CAFOS automatically and online to maintain
reliability of prediction under drifting process conditions. This invention deals
with a novel mechanism that addresses this essential requirement.
OBJECTS OF THE INVENTION
It is therefore an object of the invention to propose An improved casting
advance fault forecasting system to prevent breakouts in continuous casting
process, which eliminates the disadvantages of process drifts leading to delay in
fault prediction.
Another object of the invention is to propose a technique for online automated
adaptation of a casting advance fault forecasting system to prevent breakouts in
a thin steel-Slab caster under drifting conditions.
Thermocouple-based Breakout Prevention Systems (BPS) that work well in
conventional "thick" casters for various reasons fail to perform with desired
accuracy in thin slab casters. A CAFOS (Casting Advance Fault Forecasting
System) was developed that fulfils the need for an advance prediction system for
faults in thin slab casters that are attuned specifically to types of faults
characteristic of such casters and furthermore, where prediction is sufficiently in

advance to enable successful mitigation by responsive action. In particular, this
system attends to breakouts of type 'Longitudinal Facial Crack at Corner'. It
feeds online multiple casting process parameters into an Artificial Neural Network
(ANN) which delivers in real time a continuous value of 'Abnormality Index' (AI)
which is further processed in real time to generate occasional breakout
preventive alarms. However, the accuracy of this method decays with time due
to process drifts when the running conditions have become significantly different
from those under which the original ANN was trained.
SUMMARY OF THE INVENTION
Accordingly, there is provided an online process for detection of faults in thin
slab casters in a casting advance fault forecasting system under process-drafts
consideration mode to prevent undesirable break-outs in a continuous casting
procedure. The process was real time feedback of a long-term process quality
indicator generated in the CAFOS, which is inputted into a fuzzy pre-processing
block. When this index consistently stays at extreme values, it is an indicator of
process drift as well as the direction and magnitude of the drift. The output from
the fuzzy adaptation block is a "modulation factor" that is used to modulate
some of the system inputs in a manner that balances CAFOS outputs such that
the above indicator returns to normality; in other words it stabilizes the system in
the direction opposite to which it is advanced by the drifting process.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS
Figure 1 Schematic of a thin slab casting process of prior art.
Figure 2 Flow of computation in an Artificial Neural Network.

Figure 3 Block Diagram of an Online Automated Adaptive System of the
invention.
Figure 4 Flow of computation in a Fuzzy System according to the invention.
Figure 5 Fuzzy input set distribution of a first input according to the invention.
Figure 6 Fuzzy input set distribution of a second input according to the
invention.
Figure 7 Fuzzy output set distribution of the inputs of figures 5 & 6.
Figure 8 Flowchart detailing the flow of logic in the adaptive advance alarming
system according to the invention.
DETAIL DESCRIPTION OF A PREFERRED EMBODIMENT
The present invention provides a technique for online automated adaptation of a
CAFOS to prevent breakouts in a thin steel-slab caster under drifting conditions
of the underlying industrial process. This description is divided into two broad
sections, the first part provides an overview of the system CAFOS that is the
subject of an earlier patent application by the present applicant, and the second
part describes the details of this invention.
CAFOS provides real time analysis of the combination of online parameters that
accompany continuous casting, particularly thin slab continuous casting, and
consequent decision making in real time on whether a fault condition is
impending which can be mitigated by taking appropriate action.

The above functionality is split sequentially into two parts. First, an offline-
synthesized subset of the total set of process parameters is fed as input into an
Artificial Neural Network (ANN) (REF. [3]) which generates a dimensionless
number that represents the instantaneous quality of casting. This number varies
from 0 to 1 where the former denotes very good casting and the latter a
breakout. Hence the number can be termed as an instantaneous 'Abnormality
Index', i.e. AI, as it represents the degree of abnormality of casting. This
instantaneous AI can fluctuate within short time spans but remains within a
band; hence it is more appropriate to deal with the short-time moving average of
this AI, termed MAI. Second, the MAI is transformed into an "alarming index"
using certain modeling techniques that utilize the mid-term historical variation of
this MAI in the current casting sequence. The alarming index is usually 0 but
whenever it reaches I an alarm is raised and action taken.
The practical utility of an ANN comes into effect when there is some amount of
data available from a process or system that intrinsically encapsulates the
functionality of that process, and an observer would like to capture this
functionality. An ANN precisely serves that purpose. This data is used to 'train'
the ANN so that it learns this functionality, and when a new inputvector is
provided to the ANN it emulates the process to generate representative output.
The core of the training process is the iterative readjustment of the intra-
neuronal weights so that the final output generated by the ANN for a given input
vector comes as close as possible to the actual process output for the same input
vector, and this holds true for each of the training data records.
Structurally an ANN is made up of at least three layers of neurons, where the
first is the input layer, the last is the output layer, and the rest are the hidden
layers. This is shown in figures 2 & 3. Each neuron in a layer is connected, in

principle, to all neurons in its neighboring layer/s. The neuron first sums up all
inputs from its preceding layer neurons multiplied by the connecting weight
between the two neurons, and then provides a nonlinear transform to the sum.
This becomes its output which now serves as input to the neurons of the next
layer. The summing up relationship is

(2)
in the above relations Wik represents the weight of the connection between
neurons k in the previous layer and I in the current, and the nonlinear function
shown is the sigmoid function as an example. The outputs {y} from the first or
input layer are simply the process input values, and the outputs from the output
layer are the final process outputs corresponding to the given inputs. As is seen
clearly from figure 3, the flow of information is from left to right across the
layers.
The ANN used in CAFOS takes the process parameters of continuous casting as
inputs, and generates as output an abstract or 'conceptual' quantity - the
Abnormality Index. In a sense it monitors the health of casting. The Abnormality
Index is termed "abstract or conceptual" as it is not a physical quantity that can
be measured. There does not exist any process output data like "abnormality of
casting" on which the ANN can be trained. Instead, the data available is binary -
either 0 implying good casting or a 1 implying breakout. Hence by the principles

of training, the ANN is also expected to generate a binary output - either a 0 or
a1.
The problem is that the quality of the casting process is not at all binary
classifiable, in terms of either perfect or a breakout. Most of the time it is in
between, and much closer to 0 than to 1. And as mentioned there is no available
measure of the actual state of the casting process, which can be used to train
the ANN. That which is available from the process is a 0 or a 1. More
importantly, the process shifts from state 0 to state 1 continuously and almost
monotonously in the few minutes preceding a breakout, and the function of
CAFOS is to capture this shift in the process.
CAFOS captures this shift by using the third element of information on the
process, apart from a 0 or a 1. This is the fact that the shift from 0 to 1 is near
continuous and near monotonous. So the output from the ANN, which is also
either close to 0 or close to 1 reflecting its training master (real process) output,
is continuously tracked to identify periods of continuous and monotonous
increase, and then conclusion drawn if indeed a situation has been identified as
that which prevails just a few minutes before a breakout.
For achieving the above certain pre-requisites are necessary. First, instead of
using instantaneous AI, the moving-average AI or MAI is used. Second, the
continuous casting process is accompanied by more than three hundred online
process parameters. The ANN will be unable to handle that sort of input numbers
furthermore all calculations need to be in process real time. Hence, through
careful analysis that combines core domain knowledge on the continuous casting
process with statistical and heuristic methods, the number of inputs actually
provided to the ANN is cut down to simply 6 and these are listed hereunder.

Input 1 :difference between average wide face and narrow north face heat fluxes
Input 2 :difference between average wide face and narrow north face heat fluxes
Input 3 : average of narrow face heat fluxes
Input 4 : average friction from two cylinders
Input 5 : casting speed
Input 6 : rate of change of average friction.
To conclude on whether the current conditions qualify for raising alarm, the most
important factor is a steep rise in MAI. For identifying such a steep rise, first step
is to create a baseline value that represents good casting in the current
sequence. This is termed as NICC or 'Normality Index of Current Casting'. It is
extracted by taking a long-term average of the value of MAI. Next the running
MAI is compared against the NICC, and checked if a threshold value of the ratio
of these two quantities has been crossed. If the NICC is very close to zero (as
can happen very often) this threshold ratio is crossed very easily. To prevent this
occurrence a minimum alarming threshold of MAI is also defined. If both these
conditions are met a potential alarming situation is created.
A potential alarm is converted to an actual alarm after a few more checks. First,
there is a 'reaction period' after an earlier alarm has been raised and mitigating
action has been taken, and the situation is still to stabilize. An alarm is not raised
within this 'shadow window' of a previous alarm. Second a check for continuity
and monotony is made, i.e. the 'potential alarm' is allowed to continue for at
least ten seconds (the 'threat threshold') before the actual alarm is raised.
There is an alternate situation where the NICC is running at high levels.
Imposing the 'threshold ratio' on this high NICC implies pushing the requirement
on MAI to be very high, which would lead to missing genuine alarms. This

usually occurs when the sequence has started shortly and the time window for
calculating NICC is small. For eliminating this possibility an upper threshold,
called 'NICC threshold' is imposed on NICC- if that value is crossed, NICC is set
to that value.
According to the improvement proposed by the present invention, at the kernel
of the CAFOS, there is operably connected an ANN which by its nature needs to
be trained on plant data spanning months and possibly years of plant past
operation. The training is offline after which the ANN is put online as the kernel
of the CAFOS. The period of operation from where the training data is extracted
is called the 'training window' or TW.
It is known that industrial processes by their nature tend to drift either by wear
of equipment, quality of input materials, or modifications in process (including
upstream processes in a line). Operation of an ANN whose TW is not adjusted
according to the current process condition, shall no longer provide desired
accuracy as the fidelity of the past data, on which it was earlier accurately
trained, to contemporary running conditions falls with time. Hence a mechanism
is necessary for a CAFOS-like system to automatically, and online, adapt the
kernel ANN performance to changing process conditions. That is the subject of
this invention.
CAFOS provides two quality indicators of the running process, one, a short-time
indicator called AMI or Moving Abnormality Index, and a longer-term indicator
called NICC or Normality Index of Current Casting. Assuming that an accurate
instrument exists for physically measuring a hypothetical variable called the
"Abnormality Index" of casting, and let this hypothetical measured value be
termed as a "measured abnormality index" or SAI. Correspondingly the longer-

term measured indicator be termed as the SICC. Then, the degree to which the
MAI and NICC deviate from the SAI and SICC reflect firstly the inadequacy of the
prior art system, i.e. its inability to incorporate all influencing process parameters
due to various reasons, and secondly reflects the drift of process. It is the
second component of the deviation that keeps growing with time, and diverts the
NICC monotonously towards more and more extreme values with the MAI, the
shorter time variation, keeps fluctuating.
In reality, and as hereinabove stated, the SAI and SICC do not exist, but
operational knowledge exists as to when the process quality is "good", "normal",
or "bad". Sustained persistence of the NICC at levels that indicate opposite to
process operational knowledge, imply that the drift effect has become dominant
and adjustment is necessary.
The core novelty of this invention is to plough back the drift-indicating variable,
or NICC, back online into the kernel prediction engine to automatically neutralize
the effect of the drift, in magnitude and direction.
For this a fuzzy block is inserted ahead of the ANN block in the process flow
which accepts in the NICC output from the ANN block at the previous time step
as one of its current inputs, and along with other process inputs, generates a
modulation factor to modify some of the input variables of the ANN. This
modulation factor functions as an online adaptation of the total CAFOS system
the disadvantages of the phenomenon to drifts in the underlying industrial
process. The process flow is illustratively explained in figure 3.

Getting into specifics the Fuzzy Adaptive Block takes two inputs, one, the NICC
ploughed back from the ANN post-processing block, and two, the mold plate
thickness which affects inversely the heat transfer from the wide faces. The
output from the fuzzy adaptive block is the modulation factor with which the
wide face heat flux is multiplied. This quantity the 'wide face heat flux' or WFHF
is important as it directly impacts two of the six input variables of the ANN. More
importantly, it directly impacts the Abnormality Index that is the output from the
ANN; most of the time, the higher the WFHF, the higher the AI, and
consequently higher the NICC. Thus, if the mold plates are modified such that
they are now of higher thickness, WFHF will tend to be low and need to be
amplified. Thus, the fuzzy associative matrix (FAM) or the 'inference engine' is
designed such that higher plate thickness leads to higher amplification, i.e., the
output relates directly with this input.
The other input NICC as discussed above is used as a stabilization parameter, So
if NICC is running substantially high due to process drift, the AI output from the
ANN needs to be brought down. For that to happen, the WFHF has to reduce,
i.e. the higher the NICC, the lower the amplification factor. Now it needs to
effectively function as a damping factor, so we will call it as WFHF modulation
factor, or simply the modulation factor. This is the output from the fuzzy system.
Returning to the input NICC, if the value is high the fuzzy output has to be low,
and vice versa; i.e., the relationship is inverse.
The first input, mold plate thickness, is fuzzified using seven fuzzy sets
(membership functions). The second input NICC is fuzzified using five fuzzy sets.
The output, modulation factor, is fuzzified using nine fuzzy sets. These are
illustrat ed in figures 5-7. T he relationship between inputs and output is
encapsulated in the FAM which is presented in Table 1.


Other References
1. World Steel University website, continuous casting link:
http:wvvw.steeluniversity.org/content/html/eng/default.asp?catid=27&pagied=20
81271519 valid March 23, 2011.
2. B. G. Thomas, "Modelling of the Continuous Casting of Steel-Past, Present and
Future", Metallurgical and Materials Transactions B, Vol 33B, No.6, Dec 2002, pp.
795-812.
3. Kurt Hornik: Approximation Capabilities of Multilayer Feedforward Networks.
Neural Networks, vol. 4, 1991.

WE CLAIM :
1. In a continuous casting process of steel manufacturing chain, molten steel
is passed through a water-cooled substantially vertically-aligned lubricated
mold to emerge in the form of a continuous strand consisting of a solid
shell encapsulating the liquid material, the strand being cooled to
complete solidification using water sprays with the orientation of the
strand changed slowly to horizontal direction using a plurality of rollers,
before the soliditied shell finally cut into discrete slabs, a possibility arising
in the casting process in which a formative shell either in the mold or early
in the mold or early in the spray cooling zone tears cracks or breaks
leading to outflow of inner molten steel and its quick solidification on the
downstream machinery, generally known as a breakouts that are caused
by cracks in the formative shell near the corners of emergent strand, a
casting advance fault forecasting system (OAFOS) operably connected to
an Artificial Neural Network (ANN) adapted to prevent the break-outs by
raising advance alarms, the improved CAFOS is characterized in that a
Fuzzy adaptive block is incorporated ahead of the ANN in the process flow
to neutralize the prediction error due to process drifts in magnitude and
direction, the improved CAFOS is configured to :
acquire a long term indicator representing Normality Index of current
casting (NICC) from the CAFOS, measure the NICC and calculate a
deviation of this measured NICC in respect of a measured value of moving
abnormality index (MAI);
ploughing back the deviation value representing drift indicating variables
from the ANN post processing block into the fuzzy adaptive block
alongwith other associated process inputs;

- outputting from the ANN post-processing block and inputting into the
fuzzy adaptive block the data relating to mold plate thickness which
inversely affects the heat transfer from the wide faces of the mold; and
- multiplying the outputs from the fuzzy block representing modulation
factor and the wide face heat flux (WFHF) to accurately produce alarm
sufficiently in advance to prevent break-outs.

2. The improved casting advance fault forecasting system as claimed in claim
1, wherein the data relating to mold plate thickness is fuzzified using
seven fuzzy sets.
3. The improved casting advance fault forecasting system as claimed in claim
1, wherein the NICC data is fuzzified using five fuzzy sets.
4. The improved casting advance fault forecasting system as claimed in claim
1, wherein the modulation factor is fuzzified using nine fuzzy sets.
5. The improved casting advance fault forecasting system as claimed in claim
1, wherein the artificial neural network (ANN) is made up of at least three
layers of neurons wherein the first layer, the second layers are the input
and output layers respectively, the third layer being a hidden layer, and
wherein each neuron in a layer is connected to all neurons in its
neighboring layers.

6. An improved casting advance fault forecasting system to prevent
breakouts in continuous casting process as herein described and
illustrated with reference to the accompanying drawings.

ABSTRACT

A continuous casting process is an important links in the steel making chain as it
represents transition in steel processing from its liquid to solid state, and
impacting on stability, productivity and quality. A major disadvantage in the
process relates to "breakouts" where a formative solid shell encapsulating liquid
ruptures in the mold leading to drainage of the internal liquid and consequent
blockage of equipment and process. Thermocouple-based Breakout Prevention
System (BPS) that work well in conventional "thick" casters for various reasons
fail to perform with desired accuracy in thin slab casters. A CAFOS (Casting
Advance Fault Forecasting System) is also known that fulfills the need for an
advance prediction system for faults in this slab casters that are attuned
specifically to types of faults characteristic of such casters and furthermore,
where prediction is sufficiently in advance to enable successful mitigation by
responsive action. In particular, this prior art system attends to breakouts of type
'Longitudinal Facial Crack at Corner'. However, the accuracy of this system
deteriorates with time due to process drifts when the running conditions have
become significantly different from those under which the original components
for example, the artificial neural network of the CAFOS were trained. A novel
technique is provided for automated online adaptation of CAFOS to the possible
drifts in the process. This uses real time feedback of a long-term process quality
indicator generated in CAFOS, into a fuzzy pre-processing block that manipulates
the field inputs of CAFOS in a manner that neutralized the effects of drifts in the
underling industrial process.