UDK 669.1:004.032.26	ISSN 1580-2949
Original scientific article/Izvirni znanstveni članek	MTAEC9, 48(2)269(2014)
TOPMOST STEEL PRODUCTION DESIGN BASED ON THROUGH PROCESS MODELLING WITH ARTIFICIAL NEURAL NETWORKS
PROJEKTIRANJE PROIZVODNJE VRHUNSKIH JEKEL NA PODLAGI MODELIRANJA SKOZI PROCES Z UMETNIMI NEVRONSKIMI MREŽAMI
Tadej Kodelja1, Igor Grešovnik1,2, Robert Vertnik2,3, Božidar Sarler1'2,4
1Center of Excellence BIK, Solkan, Slovenia 2University of Nova Gorica, Nova Gorica, Slovenia 3Štore Steel d.o.o., Research, Štore, Slovenia 4IMT, Ljubljana, Slovenia bozidar.sarler@imt.si
Prejem rokopisa - received: 2013-06-14; sprejem za objavo - accepted for publication: 2013-07-03
Application of artificial neural networks for modeling of a complete process path in a steel production - from the scrap steel to the material properties of semi products - is presented. The described approach is introduced as an alternative to physics based through process modeling, with the advantage of lower complexity of the software and much lower computing times for calculating the influence of a specific settings of the process parameters. This new approach can be beneficially used in designing the production process. This is clearly demonstrated by estimating the influence of 34 alloying elements and process parameters of 6 process steps on 5 final mechanical properties of spring steel (elongation, tensile strength, yield stress, hardness after rolling and necking), based on 1879 recorded data sets from the production line in Štore Steel company. The ANN used is of a multilayer feedforward type with sigmoid activation function and supervised learning. An important feature of this approach is its dependence on accurate and sufficient data, acquired from the modeled process. Therefore, special care must be devoted to validation of the obtained model and error estimation. The reliability and other characteristics of the available data can vary to a great extent in real industrial practice, therefore analysis of the models is a highly customized task that has to be performed on a case to case basis. A flexible and easily extensible software base has been developed in the scope of the described work in order to adequately support research, development and practical application of this kind of models.
Keywords: steel production, mechanical properties of steel, artificial neural networks, response approximation, feed forward networks with back propagation
Predstavimo uporabo umetnih nevronskih mrež za modeliranje celotne procesne poti izdelave jeklenih polizdelkov - od rene do snovnih lastnosti polizdelkov. Opisani pristop je vpeljan kot alternativa fizikalnemu modeliranju skozi proces s prednostjo manjše kompleksnosti programske opreme ter bistveno manjšimi računskimi časi za izračun vpliva specifične nastavitve procesnih parametrov. Takšen pristop se lahko s pridom uporablja pri načrtovanju proizvodnega procesa. To je nazorno prikazano pri oceni vpliva 34 legirnih elementov in procesnih parametrov 6 procesnih korakov na 5 končnih snovnih lastnosti vzmetnega jekla (raztezek, natezna trdnost, meja tečenja, trdota po valjanju in skrček), na podlagi 1879 zabeleženih podatkovnih setov iz proizvodne linije podjetja Štore Steel. Uporabljena je usmerjena nevronska mreža s sigmasto aktivacijsko funkcijo in nadzorovanim učenjem. Pomembna značilnost tega pristopa je njegova odvisnost od pravilnih in zadostnih podatkov, pridobljenih iz procesa. Zato se je potrebno posebej posvetiti validaciji pridobljenega modela in oceni napak. Zanesljivost in ostale značilnosti razpoložljivih podatkov, pridobljenih iz realnih industrijskih procesov, se običajno zelo razlikujejo, zato je analiza takšnih modelov zelo specifična in mora biti narejena od primera do primera. V ta namen je bila izdelana fleksibilna in enostavno razširljiva programska oprema, ki omogoča primerno podporo raziskavam, razvoju in praktični uporabi tovrstnih modelov.
Ključne besede: izdelava jekla, mehanske lastnosti jekla, umetne nevronske mreže, aproksimacija odziva, nevronske mreže s povratnim razširjanjem napak
1 INTRODUCTION	across an extraordinary range of problem domains.
Examples can be found in almost all fields of industry as
Controlling the final mechanical properties of	well as in research areas that show promise for the
products or semi p.roducts is very important for steel	future.5 ANNs are already being used in steel production
production companies. This is a difficult task because	industries in modeling of blast furnace,6 continuous
there are a number of sequentially connected processes
where the output of one process is an input to the next	casting' steel rolling'7 etc. The first use of ANN in
one. Different physics based numerical models can be	modeling of the entire production path (also referred to
used to predict the outcomes, but their development can	as "through process modeling") has been demonstrated
be very complicated and time consuming.12 Artificial	for production of aluminum foil in8. Furthermore, a pre-
neural networks (ANN) based models3,4 can be used as	liminary study9 was made for complete steel production
an alternative to these physics based numerical models.	path, while in this study, additional parametric studies
Over the last years, ANNs have been successfully used	and sensitivity tests were added. The main drawback of
ANN models over physics based models is the fact that they can be used based on the specific training data only, and do not allow generalization to different production plants. Only the developed methodology is transferable.
In the present paper, we study the possibility of using ANN-based models as a comprehensive decision support tool in steel production. We explore the prediction of important mechanical properties of steel (elongation, tensile strength, flow limit, hardness and shrinkage) based on values of influential process parameters that determine the complete steel production path. The steel manufacturing process in the [tore Steel company and the respective available data were considered10 as a basis for the present study. The manufacturing process path consists of six individual processes:1112 steel making, continuous casting of steel, hydrogen removal, reheating, multiple stage rolling, and cooling on the cooling bed. Each of these processes can be independently modeled by a physics based numerical model.13-21 The state of the steel (shape, microstructure) of an individual process influences the downstream processing (subsequent processes in the process chain) and thus act as a part of input data (e.g. defining initial or boundary conditions) in the model of that process. This is schematically represented in Figure 1. In the current work we use another approach where an ANN is used to build a complete model of the whole production chain. We model the outcomes after the last process step and relate them to process parameters defining all processes involved in the production path. After the model is built, we can explore
the effect of variation of process parameters to the final material properties, e,g. by changing process parameters independently in parametric and sensitivity tests and observing model outputs.
2 MODELING SOFTWARE
A software for construction and use of ANN-based models has been developed in the scope of this work. The software was designed to match the challenges and requirements met when solving this kind of problems. In particular, it has to provide good flexibility in designing training strategies, filtering training data, verification of results, testing different network layouts, integration with other software, etc. This is crucial when approximating behavior of steel processing systems with large number of processing parameters. Data obtained from such systems is often inaccurate or even corrupted due to practical limitations in acquisition procedures. Response sampling can not be planned in advance but is accommodated to production schedules in the factory, therefore information available may be deficient in some regions of parameter space in order to obtain good response approximation and therefore verification of results plays an important role. The software platform has been elaborated in2223.
The Aforge.Net library is used as ANN framework.24 A convenient characteristics of neural networks is that approximation can be performed in two separate stages (Figure 2). In the training stage, the network is trained by using the sampled response (either measured or calculated by a numerical model). In the approximation stage, trained network is used for all subsequent calculations of approximated response at arbitrary values of input parameters. This gives the neural networks an important advantage over other modeling techniques, since the second stage if very fast as compared to the first stage. The software takes full advantage of this feature by separating these stages. This is especially
Figure 1: Steel manufacturing process modeling strategy Slika 1: Strategija modeliranja procesa izdelave
Figure 2: Approximation with neural networks: training a network with presented data pairs (top) and calculation of approximated response with trained network (bottom)
Slika 2: Aproksimacija z nevronskimi mrežami: učenje mreže na podlagi podatkovnih parov (zgoraj) in izračun aprokismiranega odziva z naučeno mrežo (spodaj)
important when performing extensive analyses of the considered process on the basis of the developed ANN models, or when incorporating the models in automatic optimization procedures.25,26
3 CONSTRUCTION OF THE ANN-BASED PROCESS MODEL
In the considered production setup from the [tore Steel company, the complete process is defined by 123 influential parameters (Table 1). There are 24 parameters defining the steel grade, 12 process parameters defining the continuous casting, 2 parameters the hydrogen removal, 4 parameters the reheating furnace, 31 parameters the rolling mill, 43 parameters the continuous rolling mill, and 7 parameters the cooling bed. On the output side, five mechanical properties of the final product are observed and represent the output values of the model (Table 2).
Table 1: Process parameters (input) Tabela 1: Procesni parametri (vhod)
Processes / properties	Number of parameters
Composition	24
Continuous casting of steel	12
Hydrogen removal	2
Billet reheating furnace	4
Rolling mill	31
Continuous rolling mill	43
Cooling bed	7
Total	123
Table 2: Material properties (output) Tabela 2: Snovne lastnosti (izhod)
	Elongation (A)
Final mechanical properties	Tensile strength (Rm)
	Yield stress (Rp0,2)
	Hardness after rolling (HB)
	Necking (Z)
rated for two billet dimensions (140 mm and 180 mm) which undergo considerably different process parameters. In addition, the data had to be filtered by applying a number of specially designed criteria in order to exclude corrupted data and overshoots. After these procedures, a total of 1879 data sets for dimension 140 mm have been prepared and used in the training procedure.
This data was randomly divided into disjoint training and verification sets. Training data was then used in training a feed forward neural network with sigmoid activation function, in which we iteratively minimize error of the model on this data by the back propagation algorithm. After the convergence was achieved, the model was validated on the verification set that was not involved in the training, in order to estimate its accuracy (Figure 3).
A number of training procedures with different ANN architectures and training parameters have been performed in order to find the best settings. Figure 4 shows convergence of maximum relative training errors for 15 different ANN settings.
Optimal settings (listed in Table 3) were identified by the convergence curve that reaches the lowest error at the end of the training procedure.
Table 3: ANN training and architecture settings
Tabela 3: Nastavitve učenja in arhitekture umetne nevronske mreže
Training parameters	
Learning rate	0.4
Momentum	0.6
Alpha value	1.0
Architecture	
Neurons in input layer	34
Neurons in 1s' hidden layer	25
Neurons in output layer	5
For construction of the models, data was manually collected from different databases representing production of the steelwork in year 2011. Data was first sepa-
Figure 3: Training and verification procedure for model construction Slika 3: Proces učenja in verifikacije pri izdelavi modela
Figure 4: Maximal relative training error convergence for the best 15 trained ANNs
Slika 4: Največja relativna napaka konvergence napake učenja za 15 najboljših umetnih nevronskih mrež
In order to build the final model, we trained the ANN with optimal architecture and training parameters from Table 3. The maximum number of epochs was set to 105. Training procedures were performed on the HP ProLiant DL 380 G7 workstation with 2 six core 3.47 GHz Intel Xenon X5690 processors (6*256 kB L2 and 12 MB L3 cache), with 24GB installed RAM. Trained neural network which gave us the best results was trained in approximately 13 hours. A remark should be given here, that training of ANN is indeed a cumbersome and CPU time consuming task, typically on the same order of a computational cost of a physics based model. However, when the ANN is trained, the use of it is typically several orders of magnitude faster than executing the physical model.
4 RESULTS
The training procedure results in the artificial intelligence model that relates the modeled output values v to the input parameters p:
v = v(p)	(1)
In the present context, v contains mechanical properties from Table 2 and p contains process parameters from Table 1.
The obtained model can be used for a detailed study of response of final mechanical properties on variation of process parameters, which gives operators a better insight into the process and can be used as a valuable decision support tool. This is endorsed by low computational times necessary to evaluate a single response once the model is built, which are around 10-3 s in our case.
For illustration, we show dependence of hardness on carbon fraction around different points in the space of model input parameters (Figure 5). We have randomly selected 5 sets of parameters (points in the parameter
space) from the training data. Then we varied the parameter of interest (in our case the carbon mass fraction), while the other parameters remained fixed. The parameter was varied from the minimum to the maximum value attained by that parameter within the training data.
It can be seen from Figure 5 that hardness generally increases with increasing carbon mass fraction, which is in line with the well-established metallurgical knowledge. This is observed for different fixed combinations of other parameters, while the precise form of the relation varies significantly with the values of other parameters of the model. Since influences of individual parameters are highly correlated, it is important for some purposes to study behavior over larger range of process settings. This facilitates to obtain a deeper insight in the process. The described approach employing ANN-based models is ideal for such purpose due to the short calculation times and exhaustiveness of information that is provided by such models.
In another illustrative example, we take a different point of view. Instead of focusing on influence of individual parameters, we try to obtain a broader picture of the comparative influence of different parameters on the observed outcomes. We first chose a set from the training data sets close to the center of the interval containing the measured data. We denote the vector of input parameters of this set by pc = Pi. We then varied one by one each component of the vector (i.e. the particular composition or process parameter) while the others were held fixed, and observed how the modeled quantities change as result of this variation. More precisely, we considered the following function of one variable:
uij(t) = vi (Pc1, Pc2,..., Pcj-1, t, Pcj.1,..., P c)
i = 1,... N v, j = 1,..., N p
(2)
where Np is the number of model parameters and Nv is the number of output quantities of the model. Each element of the parameter vector Pc was varied over the whole interval that the given parameter attained in the provided industrial data. The variations were then calculated for each output value (denoted by index i in equation (2) and for each parameter (index j in equation (2) and used as a measure of influence of the specific
Figure 5: Steel hardness after rolling as a function of the carbon mass fraction, calculated by the ANN model on five training sets Slika 5: Trdota jekla po valjanju kot funkcija masnega deleža ogljika, izračunana z umetno nevronsko mrežo na petih učnih množicah
Figure 6: Influence of process parameters on changes in elongation (A), ordered from the most influential one on the left, to the least influential one on the right
Slika 6: Vplivi procesnih parametrov na spremembe v elongaciji (A), urejeni od najbolj vplivnih na levi do najmanj vplivnih na desni
parameter. The results are shown in Figures 6 to 10 and are for each parameter calculated by:
Au y (t) = (max u ^ (t) - min u.. (t))
(3)
where max u^j (t) and min u^j (t) represents maximum and minimum influence of j-th parameter on i-th output value.
Figure 7: Influence of process parameters on changes in tensile strength (Rm)
Slika 7: Vplivi procesnih parametrov na natezno trdnost (Rm)
Figure 8: Influence of process parameters on changes in yield stress
(Rp)
Slika 8: Vplivi procesnih parametrov na mejo te~enja (Rp)
Figure 9: Influence of process parameters on changes in hardness after rolling (HB)
Slika 9: Vplivi procesnih parametrov na spremembe trdote po valjanju (HB)
Figure 10: Influence of process parameters on changes in necking (Z) Slika 10: Vplivi procesnih parametrov na spremembe vratu te~enja (Z)
Table 4 shows 3 most influential parameters for each material property. From the available parameters that we use for training the ANN (Figures 6 to 10), different elements of the composition of the material are the most influential for all five properties. Process parameters do not have major influence. The most important parameters obtained from the present ANN response are temperature of the liquid steel (Tcast) for elongation and hardness after rolling, temperature difference in the mould (Dtmould) for tensile strength, cooling water temperature in zone 1 (TZone1) for yield stress and cooling water flow rate in first spray system (Qsistem1) for necking. Obviously, the response of the model is not entirely expected. This indicates that the represented methodology should be used with care and finally judged by engineering expert knowledge. It is however true, that in the present model, several important process parameters are missing due to the lack of data acquisition in the plant (particularly for rolling), since a new rolling mill has been installed recently.
Table 4: The 3 most influential parameters for each mechanical property
Tabela 4: Trije najbolj vplivni parametri za posami~no mehansko lastnost
	Elongation (A)	Tensile strength (Rm)	Yield stress (Rp)	Hardness after rolling (HB)	Necking (Z)
1	Ni	C	C	Ti	V
2	Al	Cr	Ni	C	Si
3	Ti	Delta temperature in the mould	Mn	Ni	C
5 CONCLUSIONS
ANN have been used to model a complete production path in a steelwork. The developed methodology is essentially a black box modeling approach. Outcomes of the process can be predicted for arbitrary combination of process parameters without directly considering the physical background of the modeled process, but are instead relying on information about previous realizations of the process. As an example, a model of production line in the [tore Steel company was studied, reduced to 34 influential process parameters and with 5 observed properties of the final product. Several combinations of models that will include even less influential parameters will be studied in the future.
A significant advantage of the approach, as compared to the physics based numerical models, is much lower complexity of the model. There is no need to calibrate the model in order to compensate for physical simplifications and inaccurate knowledge of model constants, since the model is based on the realistic data gained from the actual process. Once the model is built, evaluation times are extremely short, in the order of a millisecond,
compared to several hours or even days that would be necessary for state-of-the-art physics based models of the same process. This represents a great advantage in tasks where large number of evaluations are required, such as automatic optimization of process parameters or detailed parametric studies.22 23 This kind of modeling has therefore a great potential to enable better insight and understanding of industrial processing, as well as to serve as a powerful decision support tool. This potential was indicated in the present paper by clearly presenting influence on individual process parameters on the outcomes.
The drawback of the approach is its dependence on reliable and abundant data that is sometimes hard to obtain. Great attention must be paid to estimation of accuracy of the model in imperfect conditions with regard to the available data.9 This will remain the main focus of future research, where influence of various factors on model accuracy will be studied which will eventually lead to procedures for reliable prediction of error bounds, which is crucial for industrial use. This will incorporate arrangements where controlled acquisition of training data is possible, e.g. by using physics based models. In this context, a large portion of work is devoted to building a flexible, modular and scalable software base to support such work. Finally, it should be noted, that the presented methodology stimulated more careful and complete data acquisition of the process parameters in [tore Steel company, needed for continuation of the present work and for better process repeatability as such.
Acknowledgment
The Centre of Excellence for Biosensors, Instrumentation and Process Control (COBIK) is an operation financed by the European Union, European Regional Development Fund and Republic of Slovenia, Ministry of Education, Science Culture and Sport. The financial support of COBIK, Slovenian Research Agency and [tore Steel company in the framework of the research program P2-0379 and the project L2-3651 is kindly acknowledged.
6 REFERENCES
1 B. [arler, R. Vertnik, S. Saletic, G. Manojlovic, J. Cesar, BHM Berg-
und Hüttenmännische Monatshefte, 150 (2005), 300-306
2B. [arler, R. Vertnik, A. Z. Lorbiecka, I. Vusanovic, B. Sencic, BHM Berg-und Hüttenmännische Monatshefte, (2013), 1-9, doi: 10.1007/ s00501-013-0147-7
R. Rojas, Neural Networks - A Systematic Introduction, SpringerVerlag, Berlin 1996
J. Kocijan, Modelling of dynamic systems with artificial neural networks and related methods, Založba Univerze v Novi Gorici, Nova Gorica 2007
F.	Fogelman-Soulie, P. Gallinari, Industrial Applications of Neural Networks, World Scientific Publishing Co, Great Britain 1998
J. Jimenez, J. Mochon, D. S. De Alaya, F. Obeso, ISIJ International, 52 (2004), 1935-1944
Y. Bissessur, Control Theory and Applications, 147 (2000), 633-640 [. Trcko, B. [arler, Use of artificial neural networks in predicting properties of household foil in Impol aluminum industry, M. Valant, U. Pirnat, Slovenska konferenca o materialih in tehnologijah za traj-nostni razvoj, Ajdovščina, 2009, Knjiga povzetkov, Zbornik, Založba Univerze v Novi Gorici, Nova Gorica 2009, 126-130 I. Grešovnik, T. Kodelja, R. Vertnik, B. Senčič, M. Kovačič, B. [arler, Computers, Materials & Continua, 30 (2012), 19-38 [tore Steel d.o.o., 2012, Available from World Wide Web: www.store-steel.si/DefaultE.asp
W. R. Irwing, Continuous Casting of Steel, The Institute of Materials, London 1993
J. G. Lenard, Primer on Flat Rolling, Elsevier, Amsterdam 2007 R. Vertnik, B. [arler, International Journal of Cast Metals Research, 22 (2009), 311-313
R. Vertnik, M. Založnik, B. [arler, Eng. Anal. Bound. Elem., 30 (2006), 847-855
G.	Kosec, M. Založnik, B. [arler, H. Combeau, Computers, Materials & Continua, 22 (2011), 169-195
A. Z. Lorbiecka, B. [arler, Computers, Materials & Continua, 18
(2010),	69-103
A. Z. Lorbiecka, R. Vertnik, H. Gjerkeš, G. Manojlovič, B. Senčič, J. Cesar, B. [arler, Computers, Materials & Continua, 8 (2009), 195-208
I. Kovačevic, B. [arler, Material Science Forum, 508 (2006), 579-584
U. Hanoglu, S. Islam, B. [arler, Materials Technology, 445 (2011), 545-547
M. Kovačič, B. [arler, Materials and Manufacturing Processes, 26
(2011),	464-474
M. Kovačič, B. [arler, Materials and Manufacturing Processes, 24 (2009), 369-374
I. Grešovnik, T. Kodelja, R. Vertnik, B. [arler, Applied Mechanics and Materials, 101-102 (2012), 838-841
I. Grešovnik, T. Kodelja, R. Vertnik, B. [arler, Application of artificial neural networks to improve steel production process, Proc. of the 15th Inter. Conf. on Artificial Intelligence and Soft Computing, Napoli, Italy, 2012, 249-255
Aforge.Net., 2012, Artificial Intelligence Library. Available from World Wide Web: http://www.aforgenet.com/. I. Gresovnik, A General Purpose Computational Shell for Solving Inverse and Optimisation Problems - Applications to Metal Forming Processes, Ph. D. thesis, University of Wales Swansea, U.K., 2000 I. Grešovnik, Journal of Mechanical Engineering, 9 (2007), 582-598
4
5
6
9
10
11
13
14
15
16
17
18
19
20
23
24
25
26