Determination of the Influence of Steelmaking Parameters on Surface Defects in Quarto Plates
1
Department of Materials and Metallurgy, Faculty of Natural Sciences and Engineering, University of Ljubljana, Aškerčeva 12, 1000 Ljubljana, Slovenia
2
SIJ Acroni d. o. o., Cesta Boris Kidriča 44, 4270 Jesenice, Slovenia
*
Authors to whom correspondence should be addressed.
Metals 2023, 13(3), 536; https://doi.org/10.3390/met13030536
Submission received: 30 January 2023
/
Revised: 26 February 2023
/
Accepted: 1 March 2023
/
Published: 7 March 2023
(This article belongs to the Special Issue Application of Neural Networks in Processing of Metallic Materials)
Abstract
:This work aimed to establish a relationship between the parameters affecting the steelmaking process and the quality of the quarto plates. We knew that the main causes of product defects in the plates were in the steelmaking process, so we sought to make changes to the process. All units in the steelmaking plant were equipped with sensors to control the working parameters, which were regularly stored in databases. These data are supplemented by the chemical composition of the molten steel at various stages of the process. To organise and analyse the huge amounts of data, data mining tools included in the Orange Software were used. For industrial use, the tree algorithm seems to be the most suitable, but we also used other models based on artificial intelligence. Unexpectedly, we obtained evidence of self-regulation and robustness in the steelmaking process. Another important result was that some additional parameters should be measured and analysed regularly, at least the amount of oligo-elements in the molten steel and the basicity of the final refining slag.
1. Introduction
We live in a time of rapid technological development, which also affects production by introducing artificial intelligence (AI) technologies into industrial processes [1]. This provides us with insight into data from statistical and causal points of view. AI enables the detection of errors in the technological process, the discovery of their causes and the improvement in the process [2]. Even though we need significant amounts of knowledge, experience, and time to study the data, our efforts are rewarded by improving the process.
The production of quarto plates can be roughly divided into four operations: steelmaking, plastic deformation during hot rolling, heat treatment, and pickling.
Steelmaking consists of a chain of processes in different reactors [3]. In the case of the electric arc furnace (EAF), the charge, usually scrap, must first be heated and melted down [4]. During this process, alloys are added, electrical and chemical energy is supplied, and various gases are added. The result is a melt with a unique chemical composition and temperature. Refining processes can be carried out in a range of reactors, from the ladle furnace (LF) in the less-equipped steel mills to a combination of vacuum reactors in most modern mills, e.g., argon oxygen decarburisation (AOD), vacuum degassing (VD), vacuum arc degassing (VAD), or vacuum oxygen decarburisation (VOD) [5,6,7,8,9]. The final step of the process is the solidification of the steel into a semi-finished product by ingot or continuous casting [4,10].
In continuous casting (CC), the solidified steel is roughly in the shape of the semi-finished product; in the production of quarto plates, it is in the shape of continuous cast slabs. Defects can occur throughout the whole production process in the steelmaking plant, from the melting to the cooling of the slab [11,12,13]. There are many reasons for this, ranging from the chemical composition of the melt to unsuitable or uncontrolled continuous casting parameters [12,13,14]. The defects that occur during steelmaking cannot be eliminated during hot rolling [15,16].
An often-overlooked parameter in steelmaking practice is the chemical composition of the slag. It is not regularly analysed, although it can affect the sulphur and phosphorus content of the steel, the amount, size, and type of non-metallic inclusions, and much more [9,11,17,18,19].
Nowadays, all metallurgical reactors are equipped with sensors to measure and record the timeline of the changing process parameters [1]. The results are stored in databases containing many unclassified and unrelated data [20,21,22]. Therefore, before starting the analyses, it is necessary to decide what should be used and what should be omitted from the databases. As such, correlations between causes and effects can easily and quickly be found. A very powerful tool for this task is AI [23,24,25,26,27].
Another way to use AI is to examine how changes in inputs (the independent variables) affect outputs (the dependent variables) [28,29,30,31,32,33,34,35,36,37,38,39]. For example, in the case of hot-rolled plates, the inputs are the steelmaking parameters, and the outputs are the quality of the hot-rolled plate. Therefore, to improve the technology, it would be helpful to identify the parameters and their values that have a decisive influence on the occurrence of defects in the semi-finished product [40,41,42,43].
The use of AI for solving problems with surface defects on hot-rolled products was successfully demonstrated in the work of Bombač et al. [16]. Artificial intelligence was also used to optimise the steelmaking process, which significantly impacts the product quality [19,24,29,30]. Using a decision tree algorithm for dephosphorisation in steelmaking was presented in the article by Phull et al. [24].
2. Materials and Methods
The literature shows that different defects on continuous casting semi-products can have the same cause and, vice versa, the same defect can have different origins [11,12,13,14]. Therefore, we attempt to find the roots of defects on hot-rolled plates and solve this task with the help of more or less advanced tools for data analysis.
2.1. Creation and Organisation of a Database
In our case, the technological route EAF-(AOD)-VOD-CC was used in the production of continuously cast steel slabs, which are the endpoint of the steelmaking process. Brackets for the AOD process are used, since the decarburisation of the molten steel without vacuum was performed only optionally.
All of the above-mentioned reactors are equipped with sensors that allow the measurement of various process parameters. Some of these parameters were only used for operational purposes, i.e., for the online control of the reactor, while others were collected and stored in various databases for further analysis. Therefore, despite the fact that the file does not only contain unique information, we were able to extract the most important areas.
The data that needed to be analysed were stored in six different files, i.e., databases.
- Chemical compositionThe chemical compositions of the steel melts were collected in this database. Each charge number was linked to at least one dataset from EAF, VOD, and CC. The chemical composition from AOD was included only if the reactor was used.
- Parameters of the steelmaking processA large amount of data from the steelmaking process was collected in the database, which could impact the quality of the finished hot-rolled steel plates. The information was highly varied, including the masses and temperatures of the empty and full ladles at different stages of the process, the amount of electrical energy consumed in each reactor, the gas consumption (oxygen and nitrogen), the continuous casting speed, and the schedule of the process. Additionally, many data points that should have no influence on the characteristics of the final product can be found in this database, e.g., the operator code for each reactor.
- Schedule databaseThe processes in each of the reactors in the electric steelmaking plant were divided into several steps, for example, the first steps of the EAF process: the time when the first basket of scrap was placed in the furnace, the time when the roof was closed, and the time when the electric power was turned on afterwards, among others.For all the above operations, the exact times were recorded with additional information, such as the mass of the charge and the amount of electrical energy.
- Continuous casting—plannedThis database is very narrow and contains only two useful parameters: the planned width and thickness of the continuous casting slab.
- Continuous casting—madeThe geometry and mass (actual and predicted) of each continuously cast slab have been stored in this database.
- Quality of hot-rolled platesThis table stores the type of defect on the hot-rolled plates with some additional information. In other words, it is a list of defects found during the final quality control.
At first glance, the role of continuous casting is overstated; however, slabs are the final product of the steelmaking plant. If we assume that plastic deformation has no significant influence on the quality of hot-rolled plates, the defects on the slabs become the origin of the production of hot-rolled plates of poor quality.
The number of records in each of the databases is shown in Table 1. There was a large amount of data stored in the databases, usable and unusable, so our first task was to make a rough estimate of which data should be omitted and which should be stored in the working database. The ratio of explorable data in the various files varied widely from 10 out of 10 in the case of Continuous casting—made to just over 25% in the case of chemical composition (see Table 1).
It can be seen from Table 1 that each database is structured differently. Therefore, much effort has been made to create a single database containing all relevant data. The charge number is the parameter that links all databases and, simultaneously, it is the only parameter that appears in all databases.
Without going into too much detail, it can be stated that the process of editing, sorting, and extracting data was demanding work. In fact, all irrelevant parameters, incomplete data vectors, and data vectors with illogical values were automatically eliminated, but the automation rules must be set beforehand. In addition, for classification, some numerical data were converted into categorical values, i.e., using AOD.
The variety of data was large and, if we use all available parameters, the data table becomes more wide than long. Most data consist of chemical compositions, each containing 22 chemical elements. Since more than one chemical composition was usually measured in the same reactor, it was obvious that the parameters to use in the analysis had to be carefully selected. The logical decision was to consider the last chemical composition from the EAF and the last chemical composition before continuous casting, i.e., from VOD, as relevant. In the first case, the purpose was to allow for the early modification of subsequent refining processes, and in the second, to know the final chemical composition of the molten steel before solidification.
The final database consists of 308 data vectors with 63 parameters, 6 of which have categorical values. Following a closer examination of the distribution of the parameters, the data can be divided into four main categories, as seen in Table 2.
As for the chemical composition, not only the usual alloying elements were considered, as can be seen in Table 3. It was speculated that the use of a comprehensive chemical composition could reveal the reason for the production of slabs with defects, which lead to the bad quality of hot-rolled plates.
Only one parameter among the process parameters has a categorical value: the use of the AOD process. Some of the parameters in the subset are related to the temperature of the melt in the reactors, and others to the electrical and chemical energy use. The last, common sense, most important parameters in the subset are those describing the solidification process during continuous casting, e.g., speed, length, sequence number, and slab number in each sequence.
The target values in our analyses were all related to quality and, as can be seen in Table 2, were five categorical values. Three of these were used in the present work.
- Plates with defects and perfect platesThe most important information from a technological point of view is how to produce defect-free hot-rolled plates. In our study, all plates without defects were marked as perfect, although the definition of defectiveness varies. In this case, the only criterion was the presence of a defect, and the parameter only has two values: defective and perfect.
- Group of defectsHot-rolled plates with defects were divided into three categories: with pull cracks, with cracks, and without defects/perfect. The number of plates with pull cracks and those with cracks is approximately equal.
- Type of defectIn this case, pull cracks were divided into three classes: ordinal, edge, and transverse edge pull cracks. The cracks were not divided into subgroups. As before, the classification ‘no defect’ must be considered as part of the list of defects. The number of defects in each type varies from 21 to 54. The same order of magnitude suggests that each defect type is adequately represented.
We used three databases for the prediction: the whole database and two subsets of it. The first collected the information related to production in EAF, and the second one the rest of the whole database, i.e., refining and continuous casting data. The aim of this decision was to find differences in the accuracy of the predictions made with a different number of input parameters.
2.2. Tool for the Organisation and Analysis of the Data—Orange Software
The open source data mining and machine learning software, Orange Software, was used to organise and analyse the data. It contains scripts written in C++ (machine learning and prior data-processing algorithms) and Python [44]. The modular structure of the software allows visual programming, making it a good choice for users with different programming skills. Furthermore, the variety of databases that can be used directly in the program is wide, e.g., CSV, TSV, and SQL [44].
In the pre-processing and filtering of data, tools for the automatic selection of data or data vectors with specified properties were used.
With Orange Software, several models are available for data processing, from the simplest, such as linear regression, to tools based on artificial intelligence (AI). Three models were used for our study.
- A tree algorithm with forwarding pruningFor industry, this model is perhaps the most appropriate because the influential parameter(s) can be vividly highlighted [24,25,26,43]. Therefore, this paper mainly presents the results of the model based on the tree algorithm and compares it with the neural network and Ada-Boost models.The accuracy of the tree algorithm-based model is affected by the tree depth, i.e., the number of layers/levels. For the industry, only a limited number of layers is useful. For example, with three layers, at least eight different combinations of influential parameters must be considered. In industrial practice, this means that eight different technologies must be used to produce the same semi-product, which is quite a lot [24,25,26,43].
- Neural Network (NN)The neural network model is based on the use of AI. Like our brain, it is made up of neurons that are interconnected. The capability of the NN depends on the number of hidden neurons and/or the maximum number of iterations. The NN is able to make the connection between input and output data. During the learning process, the NN acquires knowledge, i.e., different weights are assigned to the connections.
- Ada-Boost (AB) modelThis is an algorithm for statistical classification [27,41]. The essence of this method is to turn weak classifiers into a strong one [25,26]. The model AB works by systematically extracting one classifier from a set of classifiers at each iteration. The elements in the dataset are weighted according to their importance at each iteration. Initially, all elements are assigned a weight of 1, but as progress is made, higher values are assigned to the more severe cases. The process focuses on those classifiers that can help with the cases that are still misclassified [25,26,27,41].The advantages of this model are high its speed, simplicity, and non-demanding programming.
The entire database consists of 308 data vectors-charges with 58 independent and 5 dependent parameters, as we mentioned earlier. This database was divided into two sub-databases. As usual, on such occasions, 70% of the data vectors were used arbitrarily to train the models and the rest for testing the prediction accuracy.
3. Results and Discussion
The objective of our study was to find the optimal technological path for the production of defect-free hot-rolled plates. The predictions have meaning if they can prevent the production of defective plates. This can be done with a precisely prescribed technology in advance, based on the knowledge of the correct values of the influential parameters.
The necessary condition for this is a sufficiently good predictive ability of the model. Three levels of prediction were considered. The simple way is to determine whether the hot-rolled plate is defective. The more challenging one is to predict the group of defects, and the most difficult one is to predict the occurrence of a particular type of defect.
The sooner an error in the production schedule is found, the more time there is to fix it, so the predictions on databases with three extensions were examined. The first database contains the chemical composition of the melt at the end of the EAF section and the technical data of the furnace. The chemical composition before the start of continuous casting and technical data of (AOD), VOD, and CC are stored in the second database. The third, most comprehensive database is the summation of the two mentioned databases.
In the previous chapter, the models included in the Orange Software were briefly presented. In this part of the article, the results of the analyses are presented and discussed.
3.1. Prediction of Defect-Free Hot-Rolled Plates
3.1.1. Predictions with the Tree Algorithm
Table 4, Table 5, Table 6 and Table 7 show that 131 databases of defective plates and 177 databases of defect-free production were used in our study. Needless to say, the ratio between good and bad products does not represent the result of actual industrial practice.
The confusion matrices were used to evaluate the prediction accuracy. The meaning of the fields in the confusion matrix is explained in Table 4. The agreement of the predictions with the actual values is indicated with the bold font.
The confusion matrices showing the accuracy of the predictions are shown in Table 5, Table 6 and Table 7. Table 5 consists of the results of the predictions based on the smallest database, i.e., only data from EAF are presented. The influence of the secondary refinement of the molten steel and the technology of continuous casting on the prediction accuracy can be seen in Table 6. The confusion matrices, as a result of modelling the whole database, are compiled in Table 7. As can be seen from all these tables, the accuracy of all the predictions was not very good.
The error rate of the predictions is summarised in Table 8. The comparison of the predictions for the limited and the whole database shows that the accuracy is practically the same.
At first glance, these results seem surprising, but they must be considered in a broader context. Before this study, we assumed that the last process, continuous casting, must have the greatest impact on slab quality (i.e., the last semi-finished product of the steel shop), while the refining processes and the production of molten steel in the EAF are less important. However, the results shown in Table 8 prove that this is incorrect.
It does not matter which database we choose because the production of the molten steel is routine and, as such, follows the regular process. A good example of technology adaptation was the use of the AOD reactor. Although approximately 25% (80/308) of the steel melts were produced without AOD treatment, there was no noticeable effect on the chemical composition before continuous casting (Table 3) and the quality of the hot-rolled plates. As a result, the procedure in VOD-CC is strongly influenced by the inputs of the previous reactor. Therefore, the two subset databases are almost no longer independent of each other.
For the tree algorithm with one or two layers, the rate of correct prediction is less than 70%; with three layers, it is 75–80%, and with five layers, it is approximately 90%, as can be seen in Table 8.
It has already been said that in the industry, combinations of only a limited number of influencing parameters are useful. Combinations can be treated as a rule and, as such, can be included in the technology. The prediction accuracy of the tree algorithm with one or two layers does not meet the criteria to be successfully used in practice. In contrast, better predictions with the 5-layer model have no practical value in steelmaking practice.
3.1.2. Neural Network and Ada-Boost Model
The predictions made with these two models are significantly better than with the tree algorithm (Table 8). The accuracy of the predictions with the neural network is approximately 95% for limited databases and 100% for the entire database. The predictive ability of the Ada-Boost model is even better: a perfect match was achieved regardless of the database.
Unfortunately, industrial practice cannot benefit from these very good predictive capabilities when only a combination of several parameters drastically increases the quality of the hot-rolled plates.
3.2. Prediction of a Group of Defects
Notwithstanding the fact that the accuracy of the basic predictions is insufficient, we have performed further analysis of the predictive capabilities of the models used. For this purpose, we compared the predictions of the tree algorithm, the neural network, and the Ada-Boost model. The results are summarised in Table 9 and Table 10.
3.3. Prediction of the Defect Type
Even more unexpected results were obtained in predicting the type of defects on the hot-rolled plates. The tree algorithm with three layers only divided the plates into more or less two categories: plates with transverse pull cracks and defect-free plates (Table 10). With two more layers, other types of defects were included in the predictions (Table 10), but the accuracy of the predictions was not much better (Table 11). The use of the neural network and the Ada-Boost model improves the rate of successful classifications (Table 10) but not significantly, as can be seen in Table 11.
How the number of target values affects the prediction accuracy becomes clear when the results in Table 8 and Table 11 are compared. For the tree algorithm with three layers, the results are in the same range, although the increased number of target values decreases the accuracy. The tree algorithm with five layers gives a reasonably correct prediction of the origin of the defective plate, but a more accurate prediction is not possible. The separation between cracks and pull cracks can be satisfactorily performed with a neural network and Ada-Boost, but the prediction of different types of pull cracks is a challenging task, even for them.
To summarise the conclusions of the present work: it is known from the literature, e.g., [10,13,15], that one or more influencing parameters lead to the same surface defects and vice versa. An important result for the industry was that changing only one or two parameters from the collected data cannot increase the quality of hot-rolled plates. Although the existence of such a defect cannot be ruled out, other parameters, such as slag and alloy chemistry, must be included in the future.
Another finding was that some of the artificial intelligence tools can make very accurate predictions, but a large number of influencing parameters included in the model makes them useless for practical work.
4. Conclusions
This study was carried out with the database composed of data from 308 charges, specifically, 177 charges without defects, 54 charges with cracks and 77 charges with pull cracks. Unfortunately, the goal of identifying the most influential technological parameter or a small number of parameter combinations that affect the quality of hot-rolled plates was not achieved due to the constraints set in advance. Therefore, the secondary task was to determine the accuracy of the predictions of the exact type of defects.
Critical evaluation of the goals set:
- Ensuring the production of defect-free hot-rolled platesBased on the results of the tree algorithm, it was vividly demonstrated that only the combination of a large number of parameters can ensure the production of defect-free hot-rolled plates. Therefore, in our case, the tree algorithms were the most appropriate of the models used to provide the industry with useful data to increase the production yield.Although the prediction capabilities of the neural network, and especially the Ada-Boost model, were better, the results cannot be used in the industry because too many parameters were included in the prediction process.
- Achieving accurate predictions
- ○
- Prediction of successful or unsuccessful production.Neither the size nor the chronology of the databases has an overwhelming effect on the prediction accuracy. This surprising result is the effect of past processes on future ones, i.e., the databases are not independent.Using more layers in the tree algorithm leads to a higher prediction accuracy. The problem is that the desired accuracy can only be achieved with three or more layers or at least 32 different combinations of influential parameters. The predictions are excellent regardless of which of the databases and/or models is used.
- ○
- Prediction of a group of defectsThe prediction accuracy of the tree algorithm with three layers is less than 75%, but even with five layers, the group of the defect cannot be correctly predicted.As expected, the prediction of the defect-free plates was almost perfect. The separation between groups was better when using the Ada-Boost model but not enough to be satisfactory.
- ○
- Prediction of the type of defectsIt was pointless to expect a better prediction accuracy with a larger number of target values. Nevertheless, the near doubling of the number of mispredictions was a surprise.
The final verdict on the use of models with high predictive power in the present case is moderate. Yes, they can very successfully diagnose what must be changed to produce perfect hot-rolled plates. No, too many parameters need to be changed simultaneously to accomplish this task, which is not possible in reality.
Finally, it is necessary to point out an important positive result of the study. From the analyses, it appears that some important parameters were not taken into account. The first is the chemical composition of the final slag or at least its basicity. Another parameter that was not considered is the presence of harmful oligo-elements in the final chemical composition of the molten steel, i.e., lead and bismuth. On the basis of this study, some changes were proposed regarding the data collection.
Author Contributions
Conceptualization, M.K. and B.B.; methodology, M.K.; validation, M.K.; formal analysis, M.K.; investigation, M.K. and B.B.; resources, B.B.; data curation, B.B.; writing—original draft preparation, M.K.; writing—review and editing, M.K. and B.B.; visualization, M.K. All authors have read and agreed to the published version of the manuscript.
Funding
This research received no external funding.
Data Availability Statement
The original data is a trade secret and property of SIJ Acroni d. o. o. and therefore not available to the public.
Conflicts of Interest
The authors declare no conflict of interest.
References
- Zhou, D.; Xu, K.; Lv, Z.; Yang, J.; Li, M.; He, F.; Xu, G. Intelligent Manufacturing Technology in the Steel Industry of China: A Review. Sensors 2022, 22, 8194. [Google Scholar] [CrossRef]
- Colla, V.; Pietrosanti, C.; Malfa, E.; Peters, K. Environment 4.0: How digitalization and machine learning can improve the environmental footprint of the steel production processes. Matériaux Tech. 2020, 108, 507. [Google Scholar] [CrossRef]
- Steenkamp, J.D.; Du Preez, L. Introduction to the production of clean steel. J. S. Afr. Inst. Min. Metall. 2015, 115, 557–561. [Google Scholar]
- Madias, J. Electric Furnace Steelmaking. In Treatise on Process Metallurgy Industrial Processes, Part A, 1st ed.; Seetharaman, S., McLean, A., Guthrie, R., Sridhar, S., Eds.; Elsevier: Oxford, UK, 2014; pp. 271–300. ISBN 978-0-08-100000-7. [Google Scholar]
- Holappa, L. Secondary Steelmaking. In Treatise on Process Metallurgy Industrial Processes, Part A, 1st ed.; Seetharaman, S., McLean, A., Guthrie, R., Sridhar, S., Eds.; Elsevier: Oxford, UK, 2014; pp. 301–346. ISBN 978-0-08-100000-7. [Google Scholar]
- Stolte, G. Secondary Metallurgy: Fundamentals, Processes, Applications, 1st ed.; Stahleisen: Düsseldorf, Germany, 2007; p. 216. ISBN 978-3-514-00648-5. [Google Scholar]
- Ghosh, A.; Chatterjee, A. Secondary Steelmaking. In Ironmaking and Steelmaking: Theory and Practice, 1st ed.; Ghosh, A., Chatterjee, A., Eds.; PHI Learning Private Limited: New Delhi, India, 2008; pp. 362–396. ISBN 978-81-203-3289-8. [Google Scholar]
- Shamsuddin, M. (Ed.) Secondary Steelmaking. In Physical Chemistry of Metallurgical Processes, 1st ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2016; pp. 289–350. ISBN 978-1-119-07831-9. [Google Scholar]
- Dekkers, R.; Jokanovic, N.; Rombout, A.; Blanpain, B.; Wollants, P. Steel Cleanliness during Secondary Metallurgy of High-grade Quality Electric Steels. Steel Res. Int. 2005, 76, 475–480. [Google Scholar] [CrossRef]
- Louhenkilpi, S. Continuous Casting of Steel. In Treatise on Process Metallurgy Industrial Processes, Part A, 1st ed.; Seetharaman, S., McLean, A., Guthrie, R., Sridhar, S., Eds.; Elsevier: Oxford, UK, 2014; pp. 427–434. ISBN 978-0-08-100000-7. [Google Scholar]
- Guo, F.; Wang, X.; Wang, J.; Misra, R.D.K.; Shang, C. The Significance of Central Segregation of Continuously Cast Billet on Banded Microstructure and Mechanical Properties of Section Steel. Metals 2020, 10, 76. [Google Scholar] [CrossRef] [Green Version]
- Cui, H.; Zhang, K.; Wang, Z.; Chen, B.; Liu, B.; Qing, J.; Li, Z. Formation of Surface Depression during Continuous Casting of High-Al TRIP Steel. Metals 2019, 9, 204. [Google Scholar] [CrossRef] [Green Version]
- Xing, L.; Wang, M.; Guo, J.; Zhang, Z.; Zeng, F.; Chen, B.; Bao, Y. Study on Surface Longitudinal Crack Formation of Typical Hypoeutectoid Steel Produced on a Caster with Billet and Slab. Metals 2019, 9, 1269. [Google Scholar] [CrossRef] [Green Version]
- Chen, Y.-F.; Zhao, L.; Zuo, X.-T.; Tao, Q.-N.; Zhang, H.-B.; Li, H.; Wang, Q.-Q.; He, S.-P. Investigation and Minimization of Slag Spot Surface Defects in Continuous Casting of High Carbon Steel Billets through Statistical Evaluation. Metals 2020, 10, 878. [Google Scholar] [CrossRef]
- Bahrami, A.; Kiani Khouzani, M.; Mokhtari, S.A.; Zareh, S.; Yazdan Mehr, M. Root Cause Analysis of Surface Cracks in Heavy Steel Plates during the Hot Rolling Process. Metals 2019, 9, 801. [Google Scholar] [CrossRef] [Green Version]
- Bombač, D.; Terčelj, M.; Kugler, G.; Peruš, I. Amelioration of surface cracking during hot rolling of AISI D2 tool steel. Mater. Sci. Technol. 2018, 34, 1723–1736. [Google Scholar] [CrossRef]
- Shamsuddin, M. (Ed.) Metallurgical Slag. In Physical Chemistry of Metallurgical Processes, 1st ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2016; pp. 103–144. ISBN 978-1-119-07831-9. [Google Scholar]
- da Silva, V.L.F.; Wang, M.; da Silva Leitão, L.; de Aguiar Santos, E.; da Silva, A.C.S.A.; da Costa e Silva, A.L.V. Evaluating slag–metal deoxidation equilibrium in secondary steelmaking. J. Mater. Res. Technol. 2019, 8, 3453–3458. [Google Scholar] [CrossRef]
- Murua, M.; Boto, F.; Anglada, E.; Cabero, J.M.; Fernandez, L. A slag prediction model in an electric arc furnace process for special steel production. Procedia Manuf. 2021, 54, 178–183. [Google Scholar] [CrossRef]
- Larose, D.T.; Larose, C.T. Discovering Knowledge in Data: An Introduction to Data Mining, 2nd ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2014; pp. 165–208. ISBN 978-0-470-90874-7. [Google Scholar]
- Myatt, G.J.; Johnson, W.P. Making Sense of Data I: A Practical Guide to Exploratory Data Analysis and Data Mining, 1st ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2014; ISBN 978-1-118-40741-7. [Google Scholar]
- Myatt, G.J.; Johnson, W.P. Making Sense of Data II: A Practical Guide to Data Visualization, Advanced Data Mining Methods, and Applications, 2nd ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2009; ISBN 978-0-470-222805. [Google Scholar]
- Rokach, L.; Maimon, O. Classification Trees. In Data Mining and Knowledge Discovery Handbook, 2nd ed.; Maimon, O., Rokach, L., Eds.; Springer Science+Business Media: New York, NY, USA, 2010; pp. 149–174. ISBN 978-0-387-09823-4. [Google Scholar]
- Phull, J.; Egas, J.; Barui, S.; Mukherjee, S.; Chattopadhyay, K. An Application of Decision Tree-Based Twin Support Vector Machines to Classify Dephosphorization in BOF Steelmaking. Metals 2020, 10, 25. [Google Scholar] [CrossRef] [Green Version]
- Schapire, R.E. The strength of weak learnability. Mach. Learn. 1990, 5, 197–227. [Google Scholar] [CrossRef] [Green Version]
- Freund, Y.; Schapire, R.E. A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst. Sci. 1997, 55, 119–139. [Google Scholar] [CrossRef] [Green Version]
- Wang, S.; Li, H.; Zhang, Y.; Wang, C.; He, X.; Chen, D.; Yang, K. Studies on Parameters Affecting Temperature of Liquid Steel and Prediction Using Modified AdaBoost.RT Algorithm Ensemble Extreme Learning Machine. Metals 2022, 12, 2028. [Google Scholar] [CrossRef]
- Peruš, I.; Kugler, G.; Malej, S.; Terčelj, M. Contour maps for simultaneous increase in yield strength and elongation of hot extruded aluminum alloy 6082. Metals 2022, 12, 461. [Google Scholar] [CrossRef]
- Haruni, A.M.O.; Negnevitsky, M.; Haque, M.E.; Muttaqi, K.M. Implementation of Artificial Intelligence Technique to Model Arc Furnace Responses. In Proceedings of the Australasian Universities Power Engineering Conference (AUPEC’08), Sydney, Australia, 14–17 December 2008; pp. 1–6. [Google Scholar]
- Moghadasian, M.; Alenasser, E. Modelling and Artificial Intelligence-Based Control of Electrode System for an Electric Arc Furnace. J. Electromagn. Anal. Appl. 2011, 3, 47–55. [Google Scholar] [CrossRef] [Green Version]
- Peruš, I.; Terčelj, M.; Kugler, G. Determination of scrap/supply probability curves for the mechanical properties of aluminium alloys in hot extrusion using a neural network-like approach. Expert Syst. Appl. 2012, 39, 5634–5640. [Google Scholar] [CrossRef]
- Zou, Y.; Yang, L.; Li, B.; Yan, Z.; Li, Z.; Wang, S.; Guo, Y. Prediction Model of End-Point Phosphorus Content in EAF Steelmaking Based on BP Neural Network with Periodical Data Optimization. Metals 2022, 12, 1519. [Google Scholar] [CrossRef]
- Laha, D.; Ren, Y.; Suganthan, P.N. Modeling of steelmaking process with effective machine learning techniques. Expert Syst. Appl. 2015, 42, 4687–4696. [Google Scholar] [CrossRef]
- Terčelj, M.; Fazarinc, M.; Kugler, G.; Peruš, I. Influence of the chemical composition and process parameters on the mechanical properties of an extruded aluminium alloy for highly loaded structural parts. Constr. Build. Mater. 2013, 44, 781–791. [Google Scholar] [CrossRef]
- Ojeda Roldán, Á.; Gassner, G.; Schlautmann, M.; Acevedo Galicia, L.E.; Andreiana, D.S.; Heiskanen, M.; Leyva Guerrero, C.; Dorado Navas, F.; del Real Torres, A. Optimisation of Operator Support Systems through Artificial Intelligence for the Cast Steel Industry: A Case for Optimisation of the Oxygen Blowing Process Based on Machine Learning Algorithms. J. Manuf. Mater. Process 2022, 6, 34. [Google Scholar] [CrossRef]
- Večko Pirtovšek, T.; Peruš, I.; Kugler, G.; Terčelj, M. Towards improved reliability of the analysis of factors influencing the properties on steel in industrial practice. ISIJ Int. 2009, 49, 395–401. [Google Scholar] [CrossRef] [Green Version]
- Saparrat, M.; Monti, F.; Ibarra, H. AI Application to Melting Temperature Prediction in an Electric Arc Furnace. Iron Steel Technol. 2021, 1, 40–45. [Google Scholar]
- Carlsson, L.S.; Samuelsson, P.B.; Jönsson, P.G. Using Statistical Modeling to Predict the Electrical Energy Consumption of an Electric Arc Furnace Producing Stainless Steel. Metals 2020, 10, 36. [Google Scholar] [CrossRef] [Green Version]
- Tomažič, S.; Andonovski, G.; Škrjanc, I.; Logar, V. Data-Driven Modelling and Optimization of Energy Consumption in EAF. Metals 2022, 12, 816. [Google Scholar] [CrossRef]
- Peruš, I.; Palkowski, H.; Kugler, G.; Terčelj, M. Quantifying complex influences of chemical composition and soaking conditions for increasing the hot workability of M2 high-speed steel by using the alternative approach. J. Mater. Res. Technol. 2020, 9, 13301–13311. [Google Scholar] [CrossRef]
- Chengsheng, T.; Huacheng, L.; Bing, X. AdaBoost typical Algorithm and its application research. In Proceedings of the 3rd International Conference on Mechanical, Electronic and Information Technology Engineering (ICMITE 2017), Chengdu, China, 16–17 December 2017; pp. 1–6. [Google Scholar] [CrossRef]
- Sebastiani, P.; Abad, M.M.; Ramoni, M.F. Bayesian Network. In Data Mining and Knowledge Discovery Handbook, 2nd ed.; Maimon, O., Rokach, L., Eds.; Springer Science+Business Media: New York, NY, USA, 2010; pp. 175–208. ISBN 978-0-387-09823-4. [Google Scholar]
- Kordos, M.; Blachnik, M.; Perzyk, M.; Kozlowski, J.; Bystrzycki, O.; Gródek, M.; Byrdziak, A.; Motyka, Z. A Hybrid System with Regression Trees in Steel-Making Process. In Proceedings of the 6th International Conference on Hybrid Artificial Intelligence Systems (HAIS 2011), Wroclaw, Poland, 23–25 May 2011; pp. 222–230. [Google Scholar] [CrossRef]
- Demšar, J.; Curk, T.; Erjavec, A.; Gorup, C.; Hočevar, T.; Milutinovič, M.; Možina, M.; Polajnar, M.; Toplak, M.; Starič, A.; et al. Orange: Data Mining Toolbox in Python. J. Mach. Learn. Res. 2013, 14, 2349–2353. [Google Scholar]
Table 1.
The number of data vectors and the records they contain for each database.
| Data Vectors | Records All | Records Used | |
|---|---|---|---|
| Chemical composition | 3284 | 83 | 23 |
| Parameters of the steelmaking process | 422 | 58 | 37 |
| Schedule database | 78,128 | 14 | 5 |
| Continuous casting—plan | 422 | 6 | 3 |
| Continuous casting—made | 5388 | 10 | 10 |
| Quality of hot-rolled plates | 1104 | 11 | 10 |
Table 2.
Subsets of the database.
| Chemical Composition after EAF | Chemical Composition before CC | Process Parameters | Quality of Plate | |
|---|---|---|---|---|
| Number of Instances (Numeric/Categorical) | 22 22/0 | 22 22/0 | 14 13/1 | 5 0/5 |
Table 3.
Chemical composition of the steel melt, average value, and standard deviation before the continuous casting.
Table 3.
Chemical composition of the steel melt, average value, and standard deviation before the continuous casting.
| AOD | Without AOD | |||
|---|---|---|---|---|
| Avg. Value | ± | Avg. Value | ± | |
| C | 0.0353900 | 0.0087800 | 0.0406710 | 0.0074000 |
| Si | 0.3758800 | 0.0395400 | 0.3982290 | 0.0624700 |
| Mn | 1.8096490 | 0.0449100 | 1.8232880 | 0.0393700 |
| S | 0.0005830 | 0.0003400 | 0.0008060 | 0.0004600 |
| P | 0.0355130 | 0.0028900 | 0.0362300 | 0.0022900 |
| Cr | 17.0417000 | 0.1110000 | 17.0891000 | 0.1460000 |
| Ni | 10.6641800 | 0.1114000 | 10.6528300 | 0.0808300 |
| Cu | 0.3398330 | 0.0502600 | 0.3344690 | 0.0357300 |
| Al | 0.0181771 | 0.0075800 | 0.0196440 | 0.0080200 |
| Sn | 0.0092860 | 0.0017700 | 0.0094610 | 0.0013000 |
| Mo | 2.1108950 | 0.0393500 | 2.1178810 | 0.0504300 |
| V | 0.7768600 | 0.0126100 | 0.0749440 | 0.0126100 |
| Ti | 0.3095740 | 0.0617700 | 0.3246690 | 0.0555300 |
| Nb | 0.0203930 | 0.0128100 | 0.0116070 | 0.0046700 |
| W | 0.0784740 | 0.0330300 | 0.0711250 | 0.0218000 |
| Zr | 0.0041680 | 0.0010100 | 0.0041000 | 0.0008000 |
| B | 0.0005200 | 0.0005700 | 0.0005440 | 0.0003700 |
| Pb | 0.0020204 | 0.0009220 | 0.0022568 | 0.0007900 |
| Ca | 0.0006670 | 0.0004600 | 0.0007020 | 0.0005800 |
| Ce | 0.0251390 | 0.0055900 | 0.2439400 | 0.0044200 |
| Ta | 0.0062280 | 0.0033100 | 0.0065410 | 0.0026700 |
| Co | 0.1999537 | 0.0513900 | 0.1943290 | 0.0427300 |
Table 4.
Explanation of the meaning of the fields in the confusion matrix.
| Prediction | ||||
|---|---|---|---|---|
| Defect | Perfect | Sum | ||
| Actual | Defect | Number of defective plates predicted to be defective. | Number of defective plates predicted to be perfect. | Total number of defective plates. |
| Perfect | Number of perfect plates predicted to be defective. | Number of perfect plates predicted to be perfect. | Total number of perfect plates. | |
| Sum | Total number of plates predicted to be defective. | Total number of plates predicted to be perfect. | Total number of plates. | |
Table 5.
The confusion matrices for the prediction of defectiveness—number of instances for the EAF database.
Table 5.
The confusion matrices for the prediction of defectiveness—number of instances for the EAF database.
| Prediction | Prediction | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 117 | 14 | 131 | Actual | Defect | 113 | 18 | 131 |
| Perfect | 99 | 78 | 177 | Perfect | 80 | 97 | 177 | ||
| Sum | 216 | 92 | 308 | Sum | 193 | 115 | 308 | ||
| (a) Tree—1 layer | (b) Tree—2 layers | ||||||||
| Prediction | Prediction | ||||||||
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 96 | 35 | 131 | Actual | Defect | 106 | 25 | 131 |
| Perfect | 42 | 135 | 177 | Perfect | 16 | 161 | 177 | ||
| Sum | 138 | 170 | 308 | Sum | 122 | 186 | 308 | ||
| (c) Tree—3 layers | (d) Tree—5 layers | ||||||||
| Prediction | Prediction | ||||||||
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 120 | 11 | 131 | Actual | Defect | 131 | 0 | 131 |
| Perfect | 10 | 167 | 177 | Perfect | 0 | 177 | 177 | ||
| Sum | 130 | 178 | 308 | Sum | 131 | 177 | 308 | ||
| (e) Neural network | (f) Ada-Boost | ||||||||
Table 6.
The confusion matrices for the prediction of defectiveness—number of instances for the VOD-CC database.
Table 6.
The confusion matrices for the prediction of defectiveness—number of instances for the VOD-CC database.
| Prediction | Prediction | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 111 | 20 | 131 | Actual | Defect | 111 | 20 | 131 |
| Perfect | 87 | 90 | 177 | Perfect | 87 | 90 | 177 | ||
| Sum | 198 | 110 | 308 | Sum | 198 | 110 | 308 | ||
| (a) Tree—1 layer | (b) Tree—2 layers | ||||||||
| Prediction | Prediction | ||||||||
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 82 | 49 | 131 | Actual | Defect | 120 | 11 | 131 |
| Perfect | 16 | 161 | 177 | Perfect | 16 | 161 | 177 | ||
| Sum | 98 | 210 | 308 | Sum | 136 | 172 | 308 | ||
| (c) Tree—3 layers | (d) Tree—5 layers | ||||||||
| Prediction | Prediction | ||||||||
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 122 | 9 | 131 | Actual | Defect | 131 | 0 | 131 |
| Perfect | 10 | 167 | 177 | Perfect | 0 | 177 | 177 | ||
| Sum | 132 | 176 | 308 | Sum | 131 | 177 | 308 | ||
| (e) Neural network | (f) Ada-Boost | ||||||||
Table 7.
The confusion matrices for the prediction of defectiveness—number of instances for the complete database.
Table 7.
The confusion matrices for the prediction of defectiveness—number of instances for the complete database.
| Prediction | Prediction | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 111 | 20 | 131 | Actual | Defect | 111 | 20 | 131 |
| Perfect | 87 | 90 | 177 | Perfect | 87 | 90 | 177 | ||
| Sum | 198 | 110 | 308 | Sum | 198 | 110 | 308 | ||
| (a) Tree—1 layer | (b) Tree—2 layers | ||||||||
| Prediction | Prediction | ||||||||
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 88 | 43 | 131 | Actual | Defect | 117 | 14 | 131 |
| Perfect | 20 | 157 | 177 | Perfect | 6 | 171 | 177 | ||
| Sum | 108 | 200 | 308 | Sum | 123 | 185 | 308 | ||
| (c) Tree—3 layers | (d) Tree—5 layers | ||||||||
| Prediction | Prediction | ||||||||
| Defect | Perfect | Sum | Defect | Perfect | Sum | ||||
| Actual | Defect | 131 | 0 | 131 | Actual | Defect | 131 | 0 | 131 |
| Perfect | 0 | 177 | 177 | Perfect | 0 | 177 | 177 | ||
| Sum | 131 | 177 | 308 | Sum | 131 | 177 | 308 | ||
| (e) Neural network | (f) Ada-Boost | ||||||||
Table 8.
Error rate of predictions for the simplest task; error/complete.
| Tree Algorithm | Neural Network | Ada-Boost | ||||
|---|---|---|---|---|---|---|
| 1 Layer | 2 Layers | 3 Layers | 5 Layers | |||
| EAF | 113/ 36.7% | 98/ 31.8% | 77/ 25.0% | 31/ 10.1% | 21/ 6.8% | 0/ 0.0% |
| VOD-CC | 107/ 34.7% | 107/ 34.7% | 65/ 21.1% | 27/ 8.8% | 19/ 6.2% | 0/ 0.0% |
| EAF-VOD-CC | 107/ 34.7% | 107/ 34.7% | 63/ 20.5% | 20/ 6.5% | 0/ 0.0% | 0/ 0.0% |
Table 9.
Model capabilities for classifying defects into groups.
| Prediction | Prediction | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Cracks | Pull Cracks | Perfect | Sum | Cracks | Pull Cracks | Perfect | Sum | ||||
| Actual | Cracks | 28 | 6 | 20 | 54 | Actual | Cracks | 26 | 10 | 10 | 54 |
| Pull Cracks | 20 | 13 | 44 | 77 | Pull Cracks | 10 | 54 | 13 | 77 | ||
| Perfect | 1 | 0 | 176 | 177 | Perfect | 1 | 4 | 172 | 177 | ||
| Sum | 49 | 19 | 240 | 308 | Sum | 37 | 76 | 195 | 308 | ||
| (a) Tree—3 layers | (b) Tree—5 layers | ||||||||||
| Prediction | Prediction | ||||||||||
| Cracks | Pull Cracks | Perfect | Sum | Cracks | Pull Cracks | Perfect | Sum | ||||
| Actual | Cracks | 31 | 21 | 2 | 54 | Actual | Cracks | 39 | 15 | 0 | 54 |
| Pull Cracks | 5 | 72 | 0 | 77 | Pull Cracks | 11 | 66 | 0 | 77 | ||
| Perfect | 0 | 0 | 177 | 177 | Perfect | 0 | 0 | 177 | 177 | ||
| Sum | 36 | 93 | 179 | 308 | Sum | 50 | 81 | 177 | 308 | ||
| (c) Neural network | (d) Ada-Boost | ||||||||||
Table 10.
Model capabilities for classifying defects into types. 1 Edge transverse.
| (a) Prediction—Tree—3 layers | |||||||
|---|---|---|---|---|---|---|---|
| Cracks | Pull Cracks | Edge Pull Cracks | ET 1 Pull Cracks | Perfect | Sum | ||
| Actual | Cracks | 2 | 0 | 0 | 8 | 11 | 21 |
| Pull Cracks | 0 | 0 | 0 | 10 | 13 | 23 | |
| Edge Pull Cracks | 2 | 0 | 0 | 15 | 16 | 33 | |
| ET 1 Pull Cracks | 2 | 0 | 0 | 34 | 18 | 54 | |
| Perfect | 0 | 0 | 0 | 3 | 174 | 177 | |
| Sum | 6 | 0 | 0 | 70 | 232 | 308 | |
| (b) Prediction—Tree—5 layers | |||||||
| Cracks | Pull Cracks | Edge Pull Cracks | ET 1 Pull Cracks | Perfect | Sum | ||
| Actual | Cracks | 10 | 0 | 7 | 0 | 4 | 21 |
| Pull Cracks | 3 | 3 | 7 | 5 | 5 | 23 | |
| Edge Pull Cracks | 8 | 3 | 13 | 3 | 6 | 33 | |
| ET 1 Pull Cracks | 10 | 2 | 8 | 23 | 11 | 54 | |
| Perfect | 0 | 0 | 3 | 0 | 174 | 177 | |
| Sum | 31 | 8 | 38 | 31 | 200 | 308 | |
| (c) Prediction—Neural network | |||||||
| Cracks | Pull Cracks | Edge Pull Cracks | ET 1 Pull Cracks | Perfect | Sum | ||
| Actual | Cracks | 6 | 3 | 7 | 5 | 0 | 21 |
| Pull Cracks | 2 | 8 | 5 | 8 | 0 | 23 | |
| Edge Pull Cracks | 4 | 3 | 15 | 11 | 0 | 33 | |
| ET 1 Pull Cracks | 0 | 2 | 8 | 43 | 1 | 54 | |
| Perfect | 0 | 0 | 0 | 0 | 177 | 177 | |
| Sum | 12 | 16 | 35 | 67 | 178 | 308 | |
| (d) Prediction—Ada-Boost | |||||||
| Cracks | Pull Cracks | Edge Pull Cracks | ET 1 Pull Cracks | Perfect | Sum | ||
| Actual | Cracks | 11 | 2 | 7 | 1 | 0 | 21 |
| Pull Cracks | 2 | 14 | 6 | 1 | 0 | 23 | |
| Edge Pull Cracks | 7 | 6 | 17 | 3 | 0 | 33 | |
| ET 1 Pull Cracks | 4 | 6 | 11 | 33 | 0 | 54 | |
| Perfect | 0 | 0 | 0 | 0 | 177 | 177 | |
| Sum | 24 | 28 | 41 | 38 | 177 | 308 | |
In (a,b), it can be seen that the prediction of cracks and pull cracks with the tree algorithm is very uncertain. The results in (c,d) confirm that the neural network and Ada-Boost can correctly classify which hot-rolled plates are defect-free and which are not, but the separation between defects is insufficient.
Table 11.
The effectiveness of defect prediction; error/complete.
| Tree 3 Layers | Tree 5 Layers | Neural Network | Ada-Boost | |
|---|---|---|---|---|
| Group of defects | 91/ 26.5% | 51/ 16.6% | 28/ 9.1% | 26/ 8.4% |
| Type of defect | 98/ 31.8% | 85/ 27.6% | 59/ 19.2% | 56/ 18.2% |
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |
© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
MDPI and ACS Style
Knap, M.; Bradaškja, B. Determination of the Influence of Steelmaking Parameters on Surface Defects in Quarto Plates. Metals 2023, 13, 536. https://doi.org/10.3390/met13030536
AMA Style
Knap M, Bradaškja B. Determination of the Influence of Steelmaking Parameters on Surface Defects in Quarto Plates. Metals. 2023; 13(3):536. https://doi.org/10.3390/met13030536
Chicago/Turabian StyleKnap, Matjaž, and Boštjan Bradaškja. 2023. "Determination of the Influence of Steelmaking Parameters on Surface Defects in Quarto Plates" Metals 13, no. 3: 536. https://doi.org/10.3390/met13030536
Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.
