In the steel industry - Tata steel, India, most of the
lime produced in the lime plant is used in the steel-making process at LD shops. The quality of steel produced at LD shops depends on the quality of lime used. Moreover, the lime also helps
in the crucial dephosphorization process during steel-making.
The calcined lime produced in the lime plant goes to the laboratory for testing its final quality (CaO%), which is very difficult to
control. To predict, control and enhance the quality of lime during lime making process, five machine-learning-based models
such as multivariate linear regression, support vector machine,
decision tree, random forest and extreme gradient boosting have
been developed using different algorithms. Python has been used
as a tool to integrate the algorithms in the models. Each model
has been trained on the past 14 months’ data of process parameters, collected from level 1 sensor devices, to predict the future
quality of lime. To boost the model’s prediction performance,
hyper-parameter tuning has been performed using grid-search
algorithm. A comparative study has been done among all the
models to select a final model with the least root mean square
error in predicting and control future lime quality. After the
comparison, results show that the model incorporating support
vector machine algorithm has least value of root mean square
error of 1.23 in predicting future lime quality. In addition to this,
a self-learning approach has also been incorporated into support
vector machine model to enhance its performance further in realtime. The result shows that the performance has been boosted
from 85% strike-rate in +/-2 error range to 90% of strike-rate in
+/-1 error range in real-time. Further, the above predictive model
has been extended to build a control model which gives prescriptions as output to control the future quality of lime. For this purpose, a golden batch of good data has been fetched which has
shown the best quality of lime (≥ 94% of CaO%). A good range
of process parameters has been extracted in the form of upper
control limit and lower control limit to tune the set-points and to
give the prescriptions to the user. The integration of these two
models (Predictive model and control model) helps in controlling
the quality of lime 12 hours before its final production of lime in
lime plant. Results show that both models (Predictive model and
control model) have 90% of strike-rate within +/-1 of error in
real-time. Finally, a human machine interface has been developed to facilitate the user to take action based on control model’s
output. Eventually this work is deployed as a lime making process
automation to predict and control the lime quality.