Predicting Carbon in Steel Industries
through Machine Learning!
Predicting carbon with high accuracy in
steel steel-making Industry is a complex process. Deep learning algorithms have come
forward recently to predict carbon and sulfur accurately to avoid excessive
waste of finished products or undesired eventualities like recalling the
product from markets to avoid loss of reputation. Indeed costs are heavy if
such an event occurs.
The carbon content of the
substance not only decides its strength and its brittleness in the end product
but also affects the workability of the metal.
If an incorrectly graded piece of
steel is subsequently worked into a part, it can break because of being
too soft or too brittle which, in turn, will result in mechanical
failure. Carbon is a very important element in steel since it allows the
steel to be hardened by heat treatment. Only a small amount of carbon is
needed to produce steel: up to 0.25% for low-carbon steel, 0.25-0.50% for
medium-carbon steel, and 0.50-1.25% for high-carbon steel. Steel can contain up
to 2% carbon, but over that amount, it is considered to be cast iron, in which
the excess carbon forms graphite.
Process parameters are
dependent on the analysis of the input raw material. Since the RM used in steelmaking, does not have uniform chemistry, its variables control the process
and determine the quality of the end product.
Normally random checks are conducted
and the process is decided for each batch which becomes problematic to get
desired output. It requires control of the temperature of molten mass, heat
time, oxygen rate and other parameters depending on the impurities and the
variables present in the RM.
At
present, the regression model is considered to be the most useful which enables
predicting future responses with small variations. Without testing the samples
randomly in the lab, spectrometric constant checks and based on analysis
deciding process parameters beforehand yield a desirable quality of the end
product.
It requires a quality algorithm
considering sets and subsets of the variables in Raw materials. If scrap, iron
ore, or sponge iron is to be used as raw material in steel making, the
percentage of impurities varies from batch to batch on a big scale which
becomes problematic to control them to the desired level in the furnaces.
The melting period is the heart of Electric Arc Furnace
(EAF) operations. The EAF has evolved into a highly efficient melting. Melting
is accomplished by supplying energy to the furnace interior. This energy can be
electrical or chemical. Electrical energy is supplied via the graphite
electrodes and is usually the largest contributor in melting operations.
Refining
operations in the electric arc furnace have traditionally involved the removal
of phosphorus, sulfur, aluminum, silicon, manganese and carbon from the steel.
In recent times, dissolved gases, especially hydrogen and nitrogen, have been
recognized as a major concern. Control of the metallic properties in the batch
is important as it determines the properties of the final product. Oxygen
reacts with aluminum, silicon and manganese to form metallic oxides, which are
slag components. These metallics tend to react with oxygen before the carbon. They
will also react with FeO resulting in a recovery of iron units to every batch.
The
complexity of the steel-making process and interactive interdependency of the
various elements becomes a challenge for devising a Deep Machine Learning
model. The backbone Neuron Network has to be used diligently when the ML model
has to be implemented to get the best advantage of ML.
Since the measurement
of hot metal composition offline is not of much help, a technique for measuring
these variables with a Soft Sensor based on Neural Networks provides
optimum results. The process and output variables that include quantity and
slag as well as their composition with respect to all the important
constituents need to be trained. These process variables can be measured online
and hence the soft sensor can be used to predict the output parameters.
A supervisory control system based on
the neural network estimator and an expert system has been found to
substantially improve the hot metal quality with respect to impurities such as
excessive or undesired elements like carbon, silicon and sulfur.
The success of the Machine learning model depends on the
quality and representativeness of data used for training. It’s crucial to
ensure that the training dataset is diverse, covering a wide range of Steel
compositions and associated carbon content values. Steel making industry is energy intensive and
traditionally the efficiency of the furnaces ranges from 60 to 70% at the
maximum. An increase in productivity by applying modern
techniques can not only enhance production but can save on energy costs and can
balance the input-output ratio to the optimum perfection.
Hence,
all leading steel manufacturers are now tending to move towards ML in the
production processes. However, lack of domain knowledge while creating a most
suitable ML Model or using the most suitable algorithm, without making it
complex and without asking for huge and unreliable data while setting up the
parameters to predict and control the process parameters for quality output
optimization becomes a challenge as there always is the shortage of experts.
The steel industry has much more to
benefit from intelligent ML models than what are being applied now.
-Sanjay Sonawani
