p area of research in steel plants. In this

p { margin-bottom: 0.25cm; line-height: 120%; }

Abstract
Certain
mechanical properties of steel, i.e. elongation percentage (EL),
least yield strength (LYS) and ultimate tensile strength (UTS), form
the basis for classification of steel coils into various categories.
Methods to improve the prediction rate of such properties, using
varied chemical and physical process parameters, has remained an
integral area of research in steel plants. In this paper, a
supplementary parameter, i.e. the cooling profile of a coil in the
hot strip mill, is also taken into account along with the customary
process parameters. This additional parameter allows the Deep Neural
Network (DNN) model to predict the mechanical properties of steel for
multiple segments along the entire length of the coil instead of the
established single-segment property prediction process. Analysis of
the predicted properties along the length of the coil furnishes
information about the homogeneity of the coil cooling process and
whether the end product is suitable to be dispatched to the consumer.
Knowledge from such a model is deemed to be a useful amenity for
material science practitioners in their quest for generating improved
quality of steel.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

Keywords:
Run out table (ROT); deep neural
network (DNN); prediction and modelling; data analyses; percentage
elongation (EL); ultimate tensile strength (UTS); least yield
strength (LYS); tail-end region; cooling sequence.
___________________________________________________________________________________________

Introduction
Owing
to its high tensile strength, low cost and recyclable life-span,
steel is one of the most important green materials consumed
worldwide. Its versatility makes it a crucial material for
construction and engineering purposes. According to the end product
requirements, steel products can be broadly classified into bars,
rods and sheets. While bars and rods are usually hot-rolled, steel
sheets may be a byproduct of either hot-rolled process or cold-rolled
process. However, in spite of which process is being used, the
mechanical properties of the end product must be contained within the
permissible range. For the purpose of this project, we will
concentrate on the hot-rolled steel sheets produced by the hot-strip
mill at Tata Steel.
The
hot strip mill produces thin sheets (about 2-12mm) from slabs (about
210mm) of steel using a process called hot rolling. The main steps of
the aforementioned process include – reheating of the slab, slab
descaling, roughing mill, finishing mill, controlled cooling at the
run out table (ROT) and coiling of the sheets. Slabs from the yard
are cold and dense and require to be reheated (1190-1250°C)
and soaked before they can be rolled into thin sheets. Heat soaking
the slabs allows the internal blended elements (Ti, Nb, Co, Ni among
others) to homogenize. Once the slabs are optimally heated, they
become malleable enough to break the slab cast structure and be
rolled into intermediate sheets (approx. 30mm; 1030°C)
via the roughing mill. Six sets of rollers in the finishing mill
reduce the sheet thickness to the final desired value (2-12mm). It is
evident that the finish rolling temperature (FRT), which is typically
between 880-930°C, is impacted by the
type of rolling process adopted. On the other hand, the terminal
sheet coiling temperature (CT) is chosen based on the final
characteristics of the mechanical properties of steel. Hence
patently, being the mediator between the finishing mill and the
coiling section, the ROT cooling becomes metallurgically critical in
obtaining the ultimate product (500-550°C)
.
As demonstrated by figure 1, the production grade of various
categories of steel are the outcome of the cooling pattern used at
the ROT.
The
mechanical properties of hot-rolled steel are governed by a complex
interaction of various physical and chemical process parameters. Each
physically measurable aspect of the hot rolling process plays a
significant part in the prediction of these properties. The influence
of chemical alloying elements on the properties of steel are not only
dependent on the type of element added to the steel, but also on the
amount of the alloying element that is being added. Carbon, which is
known to aid in the strength and hardability of steel, may also have
an adverse effect on the drawability of steel .
Silicon causes low galvanizing thickness at both low concentration
and high concentration but in an intermediate concentration
precipitates a thick zinc layer. A sulphur content greater than 0.05%
is known to have a detrimental effect on steel. The convoluted
non-linear relationship between these parameters makes it difficult
to efficiently control and predict the required mechanical properties
(EL, LYS, UTS) via traditional mathematical models.

Artificial Neural Networks and
the Present Model

The recent years have seen an
up-rise in the usage of machine learning algorithms for solving
various
routine issues faced by society. The steel industry and
material science researchers have remained no exception. Artificial
Neural Networks (ANN) and other Deep Learning models use distinctive
training algorithms which are able to overcome
the difficulties
faced by traditional mathematical models and successfully relate
these sundry non-linear inputs to the mechanical output properties
of steel.

An ANN (Figure 2a) is a mesh of
artificial neurons (fundamental building blocks
of the computational
model) that are inspired by the highly-flexible neurons present in
the human brain.
The entire model imitates the flow of knowledge
gathering, information processing and information
retrieval methods
observed by the human nervous system. For computational purposes, a
bias is set to
each neuron and each pair of neurons is assigned a
weight. This set of weights and biases lays the
foundation of the
neural network and defines the function which relates the input
parameters to the output
targets. After performing a linear
combination of inputs, weights and biases, each neuron may or may
not
perform a small non-linear transformation (Figure 2b).
Mathematically, the action of an artificial neuron, j,
can be
described as follows:

o
j = ?(?(w ij x i + b i))

where, o j
is the output weight-age of neuron j; x i
and b i are the input weight-age and bias
value of the
previous neuron i; and w ij
is the weight between neurons i and j. The non-linear
transformation function ?(, ) is usually an abstract
representation of the rate of action
potential firing in the
neurons. In ANNs, it is generally taken to be a simple rectification
function or a
sigmoidal transformation like sigmoid or
hyperbolic-tangent.
While several training algorithms and network
architectures are available, the Feed Forward with Back
Propagation
Algorithm is the most widely accepted ANN training algorithm. Back
Propagation supersedes its predecessor, the perceptron, in its
ability to train hidden layers thereby escaping the restricted
capabilities of single layer networks.

Steels industries from all over
the world have been trying to use online models and ANNs to improve
the efficiency of the steel making process. In Mukhopadhyay and
Sikdar ,
an on-line ROT model was developed to predict the CT of a coil and
analyze its change over the length of the coil. Soon after,
Mukhopadhyay and Iqbal
used ANNs to predict the mechanical properties of hot-rolled steel.
Considering only FRT, CT and the chemical parameters of steel, Reddy
et al.
modeled mechanical properties of low carbon hot-rolled steel. In
2017, Lalam et al.
went ahead to broaden the application by monitoring the mechanical
properties in the continuous galvanising line of cold-rolled steel
coils. A neural network based roll-load prediction model was
developed in Yang et al.
and
Agarwal and Choudhary
modeled an ensemble data mining steel fatigue strength predictor.

x

Hi!
I'm Angelica!

Would you like to get a custom essay? How about receiving a customized one?

Check it out