FIELD OF THE INVENTION
The invention relates to a batch annealing process implemented in a batch
annealing furnace to obtain increased ductility in metals and minerals. More
particularly, the invention relates to a method to control process parameters in a
batch annealing process.
BACKGROUND OF THE INVENTION
Annealing is a process in which metals, glass and other materials are treated to
render them less brittle and more workable. In the steel industry, steel is heated
in a controlled way to obtain desired properties such as increased ductility, and
further to relieve strain acquired during the process of cold rolling.
Cold-rolled steel coils are annealed in a batch annealing furnace (BAF) as shown
in Figure 1 to obtain desired properties and mechanical strength to render them
amenable to subsequent forming operations. In the BAF, the coils are stacked
inside a furnace and annealed around 40 to 50 hours in a non-chemically-
reacting atmosphere. There are three stages in the annealing cycle namely,
heating, soaking, and cooling. Initially, the coils are heated to a temperature of
around 700°C and subsequently kept at that temperature for certain duration for
soaking. In the cooling phase, the coils are cooled initially at a slower rate up to
500°C and then at a faster rate using a bypass cooling system until the hot spot
of the coils reaches 160°C [1.2].
It is necessary to maintain a desired heating profile of the coils in order to obtain
an expected improvement in properties from the annealing cycle. Further, if the
coils are over-heated, quality problems such as sticking of strips may result.
Similarly, the cooling cycle also needs to be maintained close to an optimum
cycle, an extended cooling cycle time resulting in lower productivity. Needless to
mention achieving the optimum annealing cycle in terms of temperature and
time profiles under actual production conditions is necessary to maintain the
quality and productivity of the cold rolling mill.
However, maintaining the exact heating and cooling profiles in the BAF is difficult
due to the process variability and inaccessible conditions to measure real-time
temperatures of coils to take any corrective action. Hence, one needs to rely on
real time control models to control the heating and cooling cycles in the BAF. A
number of earlier attempts were made to design a process model.
In LOI model [3] the unsteady axisymmetric heat conduction equation in coil
identified with highest targeted cold-spot temperature is solved every 5 minutes
interval, where CT temperature considered as gas temperature is used to obtain
the Neumann boundary condition on coil. Moreover this model at the start of
annealing creates a time-segment-wise table with defined start and end
temperatures of CT: if CT temperature is not met at end of a segment it extends
time of that segment and translates / forwards the defining times of successive
segments.
In RADCON model [4] the unsteady axisymmetric heat conduction equation on
all coils individually with radiation effect, is solved every 1 seconds interval to
identify the coil with highest targeted cold-spot temperature. This model does
not define a prior control trajectory of CT temperatures, it allows temperature to
rise at maximum rate for some time, and then cuts off excess heat to enter
soaking period and after about 60% time into annealing cycle, makes a
backward-forward simulation when all future input parameters have become
stable.
However, the prior art techniques have the disadvantages that they are disabled
to accurately predict the heating, soaking and cooling patterns in the BAF during
the running process to ensure an improved quality, throughput
and savings on utilities.
OBJECTS OF THE INVENTION
It is therefore an object of the invention to propose a method to control process
parameters in batch annealing furnace.
Another object of the invention to propose a method to control process
parameters in batch annealing furnace, which allows control of thermocouple
temperature trajectory across a batch annealing cycle.
A still another object of the invention to propose a method to control process
parameters in batch annealing furnace, which is implemented in an Artificial
Neural Network Time Series in combination with numerical simulation of heat
transfer equation.
SUMMARY OF THE INVENTION
This invention uses Artificial Neural Network Time Series (ANNTS) in combination
with numerical simulation of heat transfer equation in order to accurately predict
the heating, soaking and cooling patterns in the BAF. The entire annealing cycle
involving heating, soaking and cooling phases are split into time intervals of 20
minutes (say) duration. At every time interval, a target temperature of the
Control Thermocouple (CT) is arrived at using the ANNTS,
which takes as the inputs, stack characteristics and critical waypoints on the
temperature-time trajectory including at least past 5 (say) temperatures and
predicts the next temperature. The first five times steps needed to initiate the
time series can be extracted using different means, for example, using another
ANN (an ANN Direct) which simply takes the stack characteristics as input and
predicts first five sequential temperatures. Subsequently, to predict the target
temperature profiles as a function of time, the ANNTS is activated. Here, the
temperature of every time step is calculated with the help of the temperature
profiles in the previous five time steps and other stack parameters. Further, if, at
the start of any interval in the heating phase, the temperature is more than the
specified value, the heating rate is slightly reduced so that the temperature at
the end of the time-step is matched with the targeted value.
Similarly, if at end of any interval the temperature is less than the specified
value, then the time interval is increased in steps of 10 min (say) till the targeted
temperature is reached. However, while doing this, it is ensured that the heating
rate, is at the maximum value; else the heating rate is increased first and only
then a change in interval considered. In either case, the original series is
disturbed, so a new ANNTS run is made starting from tj+1 to regenerate the
remaining trajectory, where index ni" represents the current time instance.
Similarly in the cooling phase the start- and end- temperature deviations are
addressed in an exactly opposite manner, i.e. if the start temperature of an
interval is lower than specified then the cooling rate is reduced; if the end
temperature is higher than specified, then the interval is increased in size by 10
min intervals till the desired temperature is attained. As stated above, the
remaining trajectory from tj+i (where wj" is the above mentioned interval)
onwards is regenerated after the above change. Either one or two intervals
(approx. 30 mins.) before the earliest time of attaining the CST among any of
the coils, a backward-forward simulation is executed for predicting exactly the
time of attainment of CST of all coils. The forward simulation tells if there is a
need to translate the remaining trajectory forwards or backwards, based on
attainment of CST and corresponding CSt in the simulation.
As described above, time-series predictions of the targeted temperature profiles
of the CY are done with the help of ANNTS. The training data for this ANNTS is
generated innovatively by carrying out a large mass of numerical simulations by
solving unsteady heat transfer equation on varied stack configurations. The
relevant data is generated from multiple runs of approximated heat conduction
program having different values of coil and stack characteristics like density,
specific heat, outer diameter, width and radius of the coil as inputs. In all these
runs, some arbitrary CT temperature trajectories are manually specified, all lying
within an upper and lower envelope of feasible trajectories as shown in Figure 2,
following the rule of practicality of attainment and diversity within the above
mentioned constraints. For different combinations of coil and stack parameters, a
complete unsteady heat-conduction solution is performed for different selections
of trajectories, where time has been increased in steps of 1 min and all grid point
temperatures recorded for all time steps. The HST, HSt, CST, CSt and tfinai will
emerge as the output of each such combinatorial simulation. After the
generation of large number of data sets from forward simulation different sets of
data with [(pCp),y,w.OD, Tn^Tn-3+Tn-2,Tn, HSt, HSt, CST, CSt, t* ] as the input and
[Tn+i] as corresponding output is extracted from this database to create a
training, test and validation data set for the ANIMTS.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS
Figure 1: shows schematically a single stack annealing furnace.
Figure 2 : Graphically depicts CT temperature Trajectories according to the
Invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT
In the present invention, for the first time, Artificial Neural Network is employed
in combination with heat transfer equations in order to accurately predict the
heating, soaking and cooling patterns in a BAF. The precision of a control
trajectory with the help of ANN time series is coupled with that of fast numerical
simulations to generate a more accurate, more time and cost-efficient heating
control process for BAF operations. Needless to mention, it is practically
impossible to carry out highly accurate numerical computations to estimate the
temperature profile of the stack of coils in real time in BAF because of non-
feasible requirement of computational power. On the other hand, any real-time
control model requires reasonably accurate predictions to validate and predict
the process variables. The novelty of the present invention lies in the advantages
that one can derive from the complex ANN time-series predictions that are
trained with data obtained from accurate numerical computations, which are
adaptable to a real time to determine the temperature profile of the stack of coils
in BAF.
An accurate control mechanism at the BAF is necessary to ensure good quality of
annealed products and increase productivity. Hence, it is decided to employ the
ANN to generate a control model for implementation in real-time. The time series
prediction with ANN is done with the material properties and other geometrical
parameters such as number of coils, height, diameter and thickness of coils etc.
The entire annealing cycle involving heating, soaking and cooling phases are split
into time intervals of 20 min duration termed as to, tl, t2, t3.....t^i. At every time
interval, a target temperature is arrived at based on the ANN simulations. As
shown in Figure 2, a number of possible profiles exist between the upper and
lower envelopes. The ANN model is employed to predict the target temperatures
at every time interval for the given operating and geometrical parameters. This
helps to predict the target temperature profile to initiate the time series.
Subsequently, to predict the target temperature profiles as a function of time,
the next ANN named as time-series ANN or ANNTS is activated. This model
predicts the temperature of the target coil as a function of time on the basis of
the encapsulated functionality

Where Y is the radius of the coil, W is the width of the coil, *OD' denotes outer
diameter of the coil and (pCp) is the density and specific heat of the coil. Here
(pCp) is coupled together since in heat transfer equation (pCp) appears together
in this form. Here, the temperature of every time step is calculated with the help
of the temperature profiles in the previous five time steps and other geometrical
parameters. Further, if, at the start of an interval Y in the heating phase, the
temperature is more than the specified value, the heating rate would be slightly
reduced so that the temperature at the end of the time-step is matched with the
targeted value. Similarly, if at end of an interval Y, the temperature is less than
the specified value, then the time interval is increased in steps of 10 min till the
targeted temperature is reached. However, while doing this, it is ensured that
the heating rate is at the maximum value. If the heating rate was reduced in the
previous time step, then it would be increased to the maximum value before the
increase in time-step. By doing the above mentioned corrections, since the
original time series would have been disturbed, a new ANNTS run is made
starting from fc+i to regenerate the remaining trajectory. The preceding values of
temperatures at exact time instances, t1. ti-1, ti-2,ti-3, ti-4 corresponding to uniform
time step size are regenerated by interpolation, if needed. Similarly, in the
cooling phase, the start and end temperature deviations are addressed in an
exactly analogous manner to that mentioned above, i.e., if the start temperature
of an interval y was lower than the specified value, then the cooling rate is
reduced. Further, if the end temperature of y was higher than the specified
value, then the interval y was increased in size by 10 min intervals till the
desired temperature is attained. As stated above, the remaining trajectory from
tj+i onwards is regenerated after the above change.
As described above, the time-series predictions of the temperature profiles of the
coils are done with the help of ANNs. However, it is required to provide sufficient
data to train the ANN without which the predictions would not be possible. This
generation of data is done innovatively by carrying out numerical simulations by
solving unsteady heat transfer equation with different levels of approximations in
the configurations, as described below.
Offline preparation
To execute the ANNTS and also ANND, the relevant data is generated from
multiple runs of approximated heat conduction programs having different values
of (pCp), y, w, OD as inputs, where y is the heat transfer coefficient at coil
boundary. In all these runs, some arbitrary CT temperature trajectories are
manually specified, all lying within the upper and lower envelope of feasible
trajectories shown in Figure 2, following the rule of practicality of attainment and
diversity within the above mentioned constraints. For different combinations of
coil and stack parameters, complete unsteady solutions are performed for
different selections of trajectories, starting from TCT at t=0 to TCT at t-tnnai,
where t can be increased insteps of 1 min and all grid point temperatures are
recorded at each time step. The HST, HSt, CST and thnai will emerge as the
output of each such combinatorial simulation.
After the generation of large number of data sets from these forward numerical
simulations, different sets of data with [(pCp), y, w, OD, Tn-4, Tn-3, Tn-2, Tn-i, Tn,
HST. Hst, CST, CSt, tf ] + \Tn+1] (where [(pCp), y. W. OD, Tn.4/Tn-4, Tn.3/Tn-i, Tn,
HST, HSt, CST, CSt, tf ] and the output of the current simulation i.e. [Tn+i ] are
the inputs of the next ANNTS simulation) subsets will be used to create the
training , test and validation data sets for the ANN simulations. These sets of
data will be randomly extracted from the mass of data generated in the ANN-
data-generation-runs on the approximated equations. Parameters are generated
from a meticulously designed set of highly accurate simulations, such as the
correct Y as a function of *k' and W, the flow split across different channels of
the stack as a function of W, 'OD' and no. of coils, the factor of radiation on the
heat transfer etc. In order to extract these details, number of high-fidelity runs
on FLUENT code are executed in a planned manner.
Pptential superiority
1. Because of much higher accuracy of ANNTS-simulation prediction, the
need to err-on-the conservative is significantly reduced, i.e. the need
to heat more so that the cold spot is not missed is significantly
reduced thus saving time. And also because of accuracy, the need to
overcool so as to err-on the conservative is reduced, again leading to
time saving. So, in short, the "err-on-the-conservative" based excess
time taken is reduced due to higher accuracy of ANNTS-simulation
prediction.
2. Finer control over the ramp, controlled heating and soaking sectors
leads to the following advantages:
i. Finer control over coil quality,
ii. Optimum consumption of utilities.
Keywords: Batch Annealing Furnace (BAF), Artificial Neural Network (ANN),
Control Thermocouple (CT), Artificial Neural Network Direct (ANND), Artificial
Neural Network Time Series (ANNTS), Hot Spot Temperature (HST), Cold Spot
Temperature (CST), Cold Spot time (CSt), Hot Spot time (HSt).
REFERENCES
1. Soon Kyung Kim, Moon Kyung Kim and Eon Chan Jeon, 26
November 1996, "A Study on the Annealing Characteristics of
BAF for Cold Rolled Strip", KSME International Journal, Vol. 12,
No. 2, pp. 330-337, 1998.
2. Guang Chen, Mingyan Gu, 01 January, 2007, "Simulation of Steel
Coil Heat Transfer in a High Performance Hydrogen", Institute of
Energy and environment, Anhui University of Technology, Anhui,
China.
3. LOI MODEL, www.directindustry.com/.../loi.../100-hydroqen-
atmosphere-bell-type-annealinq-furnace-21762-152737.html
United States
4. RADCON MODEL, hpp://www.rad-con.com/literature.html.
WE CLAIM:
1. A computer-implemented method to control process parameters in a batch
annealing furnace, in particular to control in real time the thermocouple
temperature trajectory across the batch annealing cycle, the annealing cycle
constituting heating, soaking, cooling phases to ensure that all the coils in a
stack attain corresponding recrystallization temperature to release potential
energy and shed anisotropic properties, the method comprising: -
splitting the annealing cycle involving heating, soaking and cooling phases
into time intervals of predetermined duration;
adapting an Artificial Neural Network Time series (ANNTS) which is
enabled to determine at every time interval a target temperature of the
control thermocouple , the ANNTS being fed as input data, the data
relating to stack characteristics and critical waypoints on a temperature -
vs- time trajectory including at least five past temperature data;
adapting an Artificial Neural Network (ANN) to initiate the first five
time steps, the ANN provided with data relating to only stack
characteristics as the input, wherein the time-series-ANN (ANNTS) is
trained by mutually generating offline random realistic trajectories of
Thermocouple temperatures in annealing cycles to extract
corresponding waypoint data, through multiple numerical simulations
of known unsteady heat transfer equations which, along with coil and
stack characteristics data acts as a training data base for said time-
series-ANN (ANNTS), and
a regeneration of forward thermocouple temperature profile by
rerunning said time series-ANN (ANNTS) whenever a deviation is
detected between the initial set profile and real-time temperature
profile.
2. The method as claimed in claim 1, wherein critical waypoint data consists
of hot-spot temperature, hot-spot time, cold-spot temperature, cold-spot time.
3. The method as claimed in claim 1, wherein the temperature of a target
coil is estimated on the basis of an encapsulated functionality such as Tn-i, = /
[{(pCp) r, w, OD, Tn/ Tn-i, Tn-2/ Tn-3, "W, where:
r = radius of the coil, w = width of the coil, OD = outer diameter of
the coil, (pQ,) = density and specific heat of the coil.
4. A computer-implemented method to control process parameters in a
batch annealing furnace substantially as herein described and illustrated in the
accompanying drawings.

This invention relates to a heating control model for Batch Annealing Furnaces
used in steelmaking. The heating, soaking and cooling phases of a batch
annealing process are divided into many small time segments. An Artificial Neural
Network (ANN) based on time series predicts the forward trajectory of the
temperatures of the Control Thermocouple (CT) at each of these time instances
as a function of the stack properties and the targeted values of Cold Spot
Temperature (CST),, Cold Spot time (CSt) and a few other desired points on the
temperature-time profile. The first few values of this profile from the start are
obtained by alternate means to initiate the time series. Violation of set
temperature (as defined by the time series) at any instances in running
conditions is reciprocated by extending the time interval till the temperature is
attained, in the process requiring the time series to be regenerated by the ANN
from the altered past-profile. The time-series ANN itself is trained from a mass of
offline transient numerical heat transfer simulations performed on representative
stacks with arbitrarily-but-realistically selected profiles of CT temperatures
(identical with that of gaseous media used for heat transfer in furnace). The
performance of the invented model is expected to be superior from the
viewpoints of quality, throughput and expense on utilities