You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320395826

Fast method for slag characterization during ladle furnace steelmaking


process based on spectral reflectance

Conference Paper · October 2017

CITATIONS READS

0 754

4 authors, including:

Asier Vicente Artzai Picon


ArcelorMittal Global Research and Development Tecnalia
14 PUBLICATIONS 31 CITATIONS 83 PUBLICATIONS 2,576 CITATIONS

SEE PROFILE SEE PROFILE

Jose Antonio Arteche


Tecnalia
7 PUBLICATIONS 33 CITATIONS

SEE PROFILE

All content following this page was uploaded by Asier Vicente on 14 October 2017.

The user has requested enhancement of the downloaded file.


Asier Vicente*A, Iñaki MacayaA, Artzai PiconB, Jose Antonio ArtecheB

Fast method for slag characterization during ladle furnace


steelmaking process based on spectral reflectance

A
Arcelor Mittal Global Research and Development Basque Center; BTecnalia.

Abstract

Hyperspectral imaging reflectance analysis has proven to be successful in online characterization


applications such as material recycling [1], soil composition analysis [2], quality control [3] among
others. The measurement of a narrow spectral reflectance of specific materials allows the use of
feature extraction and regression machine learning techniques to classify the material into a specific
group or estimate some chemical parameters under controlled conditions. A method for Fast slag
composition estimation on the ladle furnace process, together with the steel composition information
from in-process steel spectrometers, would allow implementing thermo-dynamical equilibrium models
to optimize the use of steel additives to obtain a target steel grade at the optimal additive cost.

In this work, we present a fast method for slag characterization which is based on the indirect analysis
of the spectral reflectance of the slag. This method is based on a normalization procedure to remove
the specular component of the spectra, a calibration method to correct lighting conditions and a
spectral feature extraction algorithm combined with a SVr (Support vector regression) based
regression method.

A system consisting of a hyperspectral imaging system and a calibration method has been
constructed. The system has been trained with more than 600 real slag samples taken from ladle
furnace at different ArcelorMittal steel plants. In order to cover the whole slag oxidation process, three
slag samples were taken at each heat. Each sample was analysed by XRF spectroscopy and the
regression system was trained to map the values for CaO, SiO2, .S, FeO, MnO Al2O3, MgO, P2O5
obtaining composition errors below 10% on the calibrated ladle furnace oxidation process. The
estimated slag composition was used to feed a thermo-dynamical equilibrium model that, together with
the steel composition from the in-process spectrometer estimates the required additives for the
specific steel grade. This showed lower additive costs than manual additive estimation with equivalent
final steel quality.

1
I. INTRODUCTION

The main objective of steelmaking process is to produce steel which is later


processed to obtain the final manufactured products that require very particular
requirements. Steel can be obtained by two different processes in modern steel
production: iron-making route (from iron ore) and electrical steel making route (from
scrap metal).

The electric steelmaking route differs from the ironmaking route in two fundamental
aspects: the main raw material is scrap instead of iron ore and the steel melting
phenomena takes place thanks to the thermal energy released by an electric arc and
the chemical energy from the oxygen injection. Once the oxidized steel is obtained
by any of the two mentioned routes, it is transferred to the next stage of the
production process, the secondary metallurgy process. The main characteristics of
this phase are the total deoxidation of the steel, the elimination of Sulphur and the
final adjustment of the chemical composition of the steel.

During secondary metallurgy process, a mixture of oxidized materials form the slag.
Mostly, ladle treatment consists on adjusting the chemical composition of the steel.
Different additives, such as desoxidants, lime, dolomite, fluorspar and/or various
fluxes are added in the ladle furnace to perform such treatments. The produced
impurities form a slag floating on the surface of the molten metal. The slag and the
steel fractions form a thermo-dynamical system where some elements on the slag
fraction are reduced and flow into the steel fraction and some elements on the steel
fraction are oxidized and included into the slag fraction. The quantity and the type of
additives needed depends on the mass balance and the chemical balance between
the slag and steel fraction and the desired final composition of the steel.

Although laboratory spectrometers are normally used for the later analysis of the
samples and despite having available the information on the chemical composition of
steel, the visual appearance of cooled down slag is still used in the process. Thus,
the addition of ferroalloys and other additives to liquid steel still has a human
component that is prone both to the subjectivity as well as the expertise of the
worker. The addition of ferroalloys and other additives could be optimized by knowing
with greater precision the chemical composition of slag, thereby reducing the total
cost of the productive process.

2
Spectral reflectance analysis has been previously used to characterize materials
though their chemical composition on many applications such as soil composition
analysis [5], food quality analysis [6], metal recycling [1,9] and process inspection
[7][8] among others. Slag optical characterization based on the reflectance spectra
has been already proposed. However, this method does not estimate chemical
composition of the slag and only maps a slag sample with one of the predefined slag
classes which are based on their oxidation level [10].

In this paper, we present a method that analyses the spectral reflectance of ladle
furnace slag samples and estimates their chemical composition with no need of
additional sample pre-processing. A regression algorithm has been developed to
mathematically map the spectral reflectance of the slag with its actual composition.
Specifically designed normalization and calibration steps have been proposed to
allow global model training with data taken by any capture platform on the different
plants of ArcelorMittal which allows much better generalization and robustness of the
regression model. The estimated slag composition feeds a thermodynamic
equilibrium optimization model that makes use of the estimated composition and
other process data, such as the XRF steel spectrometer data, to monitor the
thermodynamical state of the process and to optimize the additive mix required to
reach the required steel quality.

II. BASICS ON HYPERSPECTRAL TECHNIQUES

The electromagnetic spectrum of an object can be defined as the distribution of


energy intensity emitted, reflected or absorbed in a range of wavelengths
determined. Thus, an incident light beam is defined by the intensity of the beam in
each of the associated wavelengths. When this beam is incident on a surface, a part
of it is absorbed, transmitted and reflected. The reflected portion depends on the
spectrum of the incident beam, the geometry of the body and the reflectance of the
object. This reflectance is determined by the molecular characteristics of the object,
being different according to the chemical compositions.

Hyperspectral imaging, also known as imaging spectroscopy, is concerned with the


measurement, analysis and interpretation of spectra acquired from a given scene (or
specific object). In simple terms, a hyperspectral image is an extended image in
which, instead of having three RGB colour channels (red, green, blue) are available

3
from a number N of channels each of which is associated to a length of wave. This is
a combination of a spectrograph and a digital camera.

As opposed to standard images observed by the human eye, hyperspectral images


contain complete spectral information of each spatial point of the image. This image
is known as hyperspectral cube (Figure 1Figure 1) and consists of a three-dimensional
matrix, in which two dimensions represent the spatial position of the point and the
third dimension represents each spectral band.

Figure 1: Example of a hyperspectral image

III. ACQUISITION SYSTEM DESCRIPTION

A complete optical acquisition system was developed for slag characterization model
validation. This system is composed by several parts:

a) A linear hyperspectral optical sensor that receives the light reflected from the
material’s surface and gets the intensity in each of the wavelengths analysed
[11]. The used hyperspectral camera is based on a spectrograph which
captures the received radiance and separates it on different wavelengths. The
optical system presented uses a spectrograph in the VNIR range (400 -
1000nm) combined with a scientific CMOS camera that provides high
sensibility to the light, mainly in the NIR range.
b) The light is provided by wide spectrum illumination lamp that assures there is
enough intensity on the wavelength range. The lighting system is based on a
halogen bulb of 2218 lumens and forms 60º with the camera axial axis to
reduce the effect of specular illumination.
c) A V-shaped slag holder tile was designed to achieve a controlled placement of
the slag rod that ensures good quality acquisition of the slag rod radiance.

4
All these components are enclosed in a box designed to ease operative and to resist
the hard environmental conditions. The developed acquisition system is shown in
Figure 2.

Figure 2: Acquisition system installed on Arcelor steel factory at Sestao

As the illumination system provide different spectral radiance along the time due to
different factors (external extra reflections, spectral signature of the bulb, etc.) an
illumination calibration and normalization method has been defined. The main
objective of it is to handle irradiance changes allowing reuse of data coming from
different factories.

IV. SLAG CHARACTERIZATION MODEL

The spectral information contained in the signal presents great variability and it is
affected by other factors such as lighting, noise among others. For this reason, there
is a need for robust algorithms to obtain the spectral information inherent to that
signal as well as statistic techniques which allow to robustly perform the necessary
regression. Figure 5 depicts the spectral processing pipeline used for characterizing
slag.

5
Figure 3: Spectral processing pipeline

The slag characterization model is based on the spectral information extracted


directly from the conventional slag samplers used in LF process. The developed
model is based on the following elements:

 Spectra processing module: This module allows extracting a group of features


(over 2.000), per hyperspectral image, that makes possible the slag sample
description more accurate. This implies illumination correction, noise signal
reduction as well as other set of signal processing algorithms.
 A regression system composes by several regressors, which each one makes
an statistical estimation of the composition of a single element (CaO, SiO2,
MgO, Al2O3, S, FeO, F, MnO, TiO2, Na2O, Cr2O3, Cl, BaO, SrO, P2O5,
K2O, ZrO2, ZnO, CuO, IB and CaO/Al2O3 ratio).
 Training module: Enables the Inference System training to obtain the different
desired parameters, ensuring that there is no statistics bias, and providing
information about the error.

The developed spectral processing technique allows associating the reflectivity of


each element (which depends on the wavelength) with numeric values associated
with their chemical composition. In order to develop this relationship it was necessary
to have a sufficiently large set of samples of a known composition which allow
associating robustly the images captured by the hyperspectral camera with specific
numerical values. Samples population required for training depends on the variability
of slag appearance. In figure 4 two slags at different LF stages are shown.

6
Figure 4: Exampes of LF slags: high oxided slag (left), low oxided slag (right)

An extensive slag dataset was generated on several ArceloMital plants. At each heat,
three slag samples were taken at different stages of the LF process accounting for a
total 100 casts with three slag samples each. Each slag sample was analysed by a
XRF spectrometer at each factory and this data was used as ground-truth
composition. The eight more abundant components were selected for developing the
slag composition regression model.

The analysed slags were divided into train and test sets (80%-20%). The train test
dataset was used for algorithm training and provided results on this paper are based
on the remaining test dataset. The system has been validated and is currently
installed in the plants which participated to those tests.

The following table summarizes the results obtained with the slag samples with the
proposed system versus laboratory analysis of same slag samples performed using
XRF (X- ray fluorescence) by Perl technique:

TABLE I
SLAG COMPOSITION ESTIMATION RESULTS

CaO SiO2 Al2O3 S Fe2O3 MnO MgO F


Max
75.61 33.07 8.87 2.80 9.84 10.02 10.53 10.60
value
Min
37.09 10.94 0.43 0.13 0.45 0.10 2.49 0.39
value
error
5.47 6.81 8.05 9.54 3.61 4.63 9.12 3.61
(%)
SD
5.19 7.61 9.55 8.14 5.39 7.16 8.26 5.39
(%)
r 0.76 0.65 0.32 0.80 0.60 0.73 0.59 0.60

It can be appreciated that errors lower than 10% are obtained on the slag
composition for the eight most abundant slag components. The regression plots
obtained for these compounds are shown in Figure 5.

7
Figure 5: Regression plots for CaO, SiO2,.S,FeO,MnO and F

V. LADLE FURNACE PROCESS OPTIMIZATION BASED ON THE NEW


TOOL

In Ladle Furnace (LF) process the oxidized liquid steel coming from EAF is
transformed into steel with very low oxygen content and a chemical composition
adjusted to the composition required in the final product. To do so, desoxidants,
fluxes and ferroalloys materials are added and by means of a stirring mixing process
between the steel and the slag with an inert gas, the desired steel conditions are
achieved, transferring all undesirable elements to the slag.

So that during the ladle process, the chemical evolution of the slag is as important as
the chemical evolution of steel. In this sense, the standard analytical technique used
for steel analysis is OES spectroscopy, however, the slag is extracted periodically
and analysed visually by an skilled operator, who makes operational decisions based
on his experience.

The thermodynamic analysis of the process allows to carry out the study of the ladle
treatment efficiency once all process data are available. This thermodynamic
equilibrium analysis gives, for a global composition of a given system, which is
defined by the quantity of each chemical element and a given temperature and
pressure, the list of phases present at equilibrium (metallic phases, slag, gas,
precipitates,…), their respective quantities (kg/t of slag, amount of inclusions, gas
volume,…) and their composition at equilibrium state.

8
There are several simulation software that can be used, based on thermodynamic
models, for doing the metallurgical process interpretation. Among them, the most
known are: "ThermoCalc" (KTH, Stockholm) [12], "FactSage" (Ecole Polytechnique,
Montréal – GTT, Aachen) [13], "MTDATA" (NPL, London) [14] or "CEQCSI"
(ArcelorMittal) [15, 16].

CEQCSI model has been used as secondary metallurgy process simulation software
to optimize the use of additives. For the steel-slag phase equilibrium model the
following variables were defined: mass of the steel fraction, mass of the slag fraction,
chemical composition of steel, chemical composition of slag, process temperature
and atmospheric pressure.

Each time the system receives a slag reflectance image, slag composition is
calculated and current network variables for steel composition are gathered from
plant network database. Thermodynamic equilibrium model is calculated for current
process state. This predicts the thermodynamic equilibrium final composition for both
steel and slag fractions. Thermodynamic equilibrium steel fraction composition is
checked against the required steel composition that depends on its quality number. If
some of the steel composition parameters are out of ranges, thermodynamic
equilibrium simulations are done by simulating addition of pre-characterized
additives. These additives are characterized both in terms of economic cost and its
real composition as a function of the thermodynamic model input parameters.

A cost function is defined to select the additive mixture that minimizes the function
cost fulfilling quality requirements.

The acquisition system is connected with the plant network to exchange data about
the whole process. Once the slag composition sample is acquired, the latest
spectrometer analysis for that heat, as well as all process available data, are
gathered from the steel spectrometer. This information was provided to the optimizer
to obtain the expected final steel composition. This composition was compared with
the steel grade requirements and an additional quantity of additives were proposed..
The system user interface displays all the information it generates and guides the
operator through the entire LF process.

9
Figure 6: System user interface: estimation of slag composition components (left) and additive optimization results
(right).

In order to analyse the goodness of the optimizations performed by the new tool in
one of the tested plant, between May 22nd and June 13th 2013, Ladle Furnace
operator followed optimizations operations in parallel with the usual treatment. As
baseline we used the process data from the operator steel treatment according to his
professional judgment looking for the quality specifications required by continuous
casting machine. In parallel to the usual treatment, the process optimization
proposals were used to evaluate the outcome of both proposals.

Table II shows a comparative analysis between the amounts of additives made by LF


operator and the proposal made by the new tool during the validation period and
extrapolating the results to an annual production scenario:

TABLE II
CALCULATED VALUE GENERATION
Price (€/kg) Reduction Reduction (%)
2011 (Kg/lst)
CaO 0,067 0,93 11,5
FeSi 1,352 0,38 14,3
FeMn 0,918 0,64 73,6

SiMn 0,926 -0,49 -4,3


CaF2 0,225 0 0
Savings (€/lst) 0.710
*Lst= Liquid steel tonnes

So that in an annual production scenario of 1 million tonnes of liquid steel, the


savings achieved would be 710 k€.

10
VI. CONCLUSIONS

The aim of this paper was to detail the theoretical and practical issues associated
with the implementation of an integrated system for slag composition characterization
and its integration into a thermodynamical equilibrium model to estimate the optimal
additives use on the ladle furnace process.

An industrial acquisition system was fully developed. This system copes with
calibration and normalization processes that assure that spectra acquired from
different acquisition sources can be used to train the same regression model.
Besides that, integration with factory level 2 network provides useful information to
perform additional actions.

The developed slag composition model has been validated. Depending on the
analysed element, the relative standard deviation between compositions and the
XRF known technique ranges between 5% and 10%. This error assures enough
precision in composition during the production process to make a quick assessment
of physical state of the slag, so as to define the steel treatment strategy to reach the
final process stage at lowest cost.

Thanks to the above described features, the method is fast and does not disrupt the
steel production process. The method provides an accurate chemical composition in
two to three seconds, whereas prior art methods require sending samples to a
laboratory for analysis.

The method makes possible to optimize the mix of additives to be added to the steel
during production in order to obtain a predetermined steel chemical composition.
Cost reduction derived of the use of the slag composition estimation model together
with thermodynamical models have been validated and quantified.

The proposed system allows monitoring the evolution of the ladle treatment at any
time (steel composition, steel and slag weights, additives, temperature…). These
data together with the information about slag composition extracted allows
conducting a global mass balance analysis at any point, so that the thermodynamic
equilibrium allow optimizing additives mixtures and reducing the use of additives.

11
VII. ACKNOWLEDGMENTS

Partial financial support of this work by the Basque Government (Etorgai NUPROSS
ER-2010/00001 and DAVOS ER-2014/0004 Projects) is gratefully acknowledged.

References

[1] Picon A, Ghita O, Bereciartua A, Echazarra J, Whelan PF, Iriondo PM. Real-time hyperspectral
processing for automatic nonferrous material sorting. Journal of Electronic Imaging. 2012;21(1):
013018–1
[2] Cavalli RM, Pascucci S, Pignatti S. Hyperspectral remote sensing data to map hazardous materials in a
rural and industrial district: The Podgorica dwellings case studies. In: 2009 First Workshop on
Hyperspectral Image and Signal Processing: Evolution in Remote Sensing. 2009. p. 1–4.
[3] Gruber F, Wollmann P, Schumm B, Grählert W, Kaskel S. Quality Control of Slot-Die Coated
Aluminum Oxide Layers for Battery Applications Using Hyperspectral Imaging. Journal of Imaging.
2016;2(2):12.
[4] http://spain.arcelormittal.com/who-we-are/management.aspx
[5] Rossel RV, Walvoort DJJ, McBratney AB, Janik LJ, Skjemstad JO. Visible, near infrared, mid infrared
or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties.
Geoderma. 2006;131(1):59–75.
[6] Cen H, He Y. Theory and application of near infrared reflectance spectroscopy in determination of food
quality. Trends in Food Science & Technology. 2007;18(2):72–83.
[7] Willoughby CT, Folkman MA, Figueroa MA. Application of hyperspectral imaging spectrometer
systems to industrial inspection. In: Photonics East’95 [Internet]. International Society for Optics and
Photonics; 1996 [cited 2017 Mar 20]. p. 264–272. Available from:
http://proceedings.spiedigitallibrary.org/proceeding.aspx?articleid=1009642.
[8] Tatzer P, Wolf M, Panner T. Industrial application for inline material sorting using hyperspectral
imaging in the NIR range. Real-Time Imaging. 2005;11(2):99–107.
[9] Picón A, Ghita O, Whelan PF, Iriondo PM. Fuzzy spectral and spatial feature integration for
classification of nonferrous materials in hyperspectral data. IEEE Transactions on Industrial
Informatics. 2009;5(4):483–494.
[10] Rodriguez S, Picon A, Gutierrez JA, Bereciartua A, Iriondo P. Automatic slag characterization based on
hyperspectral image processing. In: 2010 IEEE 15th Conference on Emerging Technologies Factory
Automation (ETFA 2010). 2010. p. 1–4.
[11] Hyvarinen TS, Herrala E, Dall’Ava A. Direct sight imaging spectrograph: a unique add-on component
brings spectral imaging to industrial applications. In: Photonics West’98 Electronic Imaging [Internet].
International Society for Optics and Photonics; 1998
[12] Andersson, J. O., Helander, T., Höglund, L., Shi, P., & Sundman, B. (2002). Thermo-Calc & DICTRA,
computational tools for materials science. Calphad, 26(2), 273-312.
[13] Bale, C. W., Chartrand, P., Degterov, S. A., Eriksson, G., Hack, K., Mahfoud, R. B., ... & Petersen, S.
(2002). FactSage thermochemical software and databases. Calphad, 26(2), 189-228.
[14] Davies, R. H., Dinsdale, A. T., Gisby, J. A., Robinson, J. A. J., & Martin, S. M. (2002). MTDATA-
thermodynamic and phase equilibrium software from the national physical laboratory. Calphad, 26(2),
229-271.
[15] Blond, E., Schmitt, N., Hild, F., Blumenfeld, P., & Poirier, J. (2007). Effect of slag impregnation on
thermal degradations in refractories. Journal of the American Ceramic Society, 90(1), 154-162.
[16] Résumé des articles techniques. (2007). Revue De Métallurgie, 104(12), III-IV.
doi:10.1051/metal/20071201

12

View publication stats

You might also like