SciELO - Scientific Electronic Library Online

 
vol.70 issue2Evaluation of hydrogen-Induced cracking resistance of the In625 laser coating system on a C-Mn steel substrateInfluence of SEBS-MA and SBS compatibilizers on properties and morphology of blends of polystyrene/rubber residue (SBRr) from the footwear industry author indexsubject indexarticles search
Home Pagealphabetic serial listing  

Services on Demand

Journal

Article

Indicators

Related links

Share


REM - International Engineering Journal

On-line version ISSN 2448-167X

REM, Int. Eng. J. vol.70 no.2 Ouro Preto Apr./June 2017

https://doi.org/10.1590/0370-44672016700032 

Metallurgy and materials

Use of an Artificial Neural Network in determination of iron ore pellet bed permeability

Marcelo Chagas1 

Marcelo Lucas Pereira Machado2 

João Batista Conti Souza3 

Eduardo F. de Jesus Frigini4 

1Professor, Instituto Federal do Espírito Santo - IFES, Departamento de Engenharia Mecânica, Cachoeiro de Itapemirim - Espírito Santo - Brasil. marcelo_chg@hotmail.com

2Professor Titular, Instituto Federal do Espirito Santo - IFES, Programa de Pós-Graduação em Metalurgia e Materiais - PROPEMM, Vitória - Espirito Santo - Brasil. marcelolucas@ifes.edu.br

3Professor, Instituto Federal do Espírito Santo - IFES, Departamento de Engenharia Metalúrgica, Vitória - Espirito Santo - Brasil. jbconti@ifes.edu.br

4Engenheiro Eletricista Sênior, Mestre em Automação e Controle, Universidade Federal do Espírito Santo - UFES, Vitória - Espírito Santo - Brasil. eduardofrigini@vale.com.br


Abstract

The thermal processing of iron ore pellets in pelletizing plants is a decisive stage regarding final product quality and knowledge of its characteristics has a fundamental importance in its process optimization. This study evaluated the variable sensitivity involved in pellet bed formations and their permeability using the artificial neural networks method. The model stated that standard diameter deviation, sphericity and pellet bed height mostly affect bed permeability. The computational model was able to predict pellet bed backpressure by means of pellet geometrical features, thus allowing improving green pellet generation, in order to ensure fuel and energy consumption reduction, final quality improvement and better productivity.

Keywords: pellet; artificial neural networks; iron ore

1. Introduction

The agglomeration of iron ore appeared with the necessity to optimize the production process of raw materials connected to the steel industry, using the fines portion generated during the mining process and subsequent ore processing, thus transforming it into an iron-rich material, which until then would be process waste, thus becoming a marketable product with a higher value (Thella and Venugopal, 2011). The iron ore pellet is a cluster of approximately sphere shaped particles with some remarkable properties as: particle size distribution between 9 mm and 15 mm, high iron content (around 66%), uniform mineralogical properties, and high mechanical strength, along with a uniform, low tendency to abrasion and good behavior during transport (Meyer, 1980). The iron ore pellet production process comprises several steps, such as the separation of raw material, grinding, classification according to particle size, agglomeration, drying processes, burning and cooling.

For traveling grate kilns, herein presented, 90% of the heat transport in thermal processing is done by convection. Thus, the gas flow through the pellet bed is of fundamental importance (Meyer, 1980).

The gas flow is dependent on the strength and gas permeability of the bed, and highly influenced by the physical consistency of the pellets. The most adopted equation to solve gas flow through packed beds is Ergun's equation (Hinkley et al., 1994; Luckos and Bunt, 2011; Trahana et al., 2014; Koekemoer and Luckos, 2015; Erdin et al., 2015):

ΔPL=A(1ε)2ε3μUDp2+B1εε3ρU2Dp

Where: ΔP - pressure drop of packed bed; L - height of packed bed; ε - void fraction; U - superficial bed velocity; µ - dynamic viscosity of the fluid; Dp - particle diameter; A and B - empirical coefficients.

The pressure variation is substantially high as there is a large variation in pellet diameters, the presence of fines, (Hollands and Sullivan, 1984) and according to Ergun's equation, varies linearly with the increase of bed height. The pressure drop is also responsible for the operation of the fans that drive the air heated by the various stages of the firing process.

Pellets with diameters distributed in a narrow band, with spherical shape and good mechanical strength, and low generation of fine particles, produce a bed in the drying step with high permeability, thus lowering production costs, improving productivity and increasing the quality of the final product (Matos, 2007).

It is therefore evident that the resistance encountered by the air flow through the pellet bed can be understood as a measurement of the quality factor of the raw pellets, providing an opportunity for energy savings in the overall process as well as a point which offers final product quality improvement.

The complexity of the equipment, the large amount of interrelated factors in the process and the difficulty of developing manufacturing condition experiments lead us to the mathematical modeling that has a broad application in metallurgy (Timofeeva et al., 2013).

The method of artificial neural networks has shown great reliability and quick responses in several areas, proving to be a suitable choice when you want to get a general predictive analysis that includes a lot of variables (Dwarapudi et al., 2007). Neural networks have the ability to capture nonlinear and highly complex relationships between inputs and outputs of the process. They are computationally efficient and do not require prior knowledge of the process to be modeled (Haykin, 2009).

2. Material and method

A sequence of 1000 industrial data collected from the pelletizing process have been used to create an artificial neural network (ANN) with the Multilayer Perceptron (MLP) architecture, feedforward type, with a seven neuron input layer composed of process variables. The data was processed with Multiple Back Propagation software (Lopes and Ribeiro, 2003). Table 1 shows the process average values, minimum, maximum and standard deviation.

Table 1 Date set description. 

RNA Variable Mean Min. Max. Std Dev. Unit
Inputs
E 1.1 Raw pellets production 699.79 291.32 753.85 53.52 ton/h
E 1.2 Diameter mean size 11.87 4.14 12.71 0.705 mm
E 1.3 Diameter standart deviation 2.32 0.77 2.64 0.147 mm
E 1.4 Pellets between 10-16 mm 76.29 26.59 80.91 4.98 %
E 1.5 Pellets above 16 mm 3.74 0.88 11.85 1.13 %
E 1.6 Pellets below 10mm 16.36 5.36 31.69 2.5 %
E 1.7 Esphericity 79.34 28.36 80.26 4.62 %
Output
S 1.1 Bed Backpressure -239.01 -282.64 -78.81 27.21 mBar

The hardware used was a regular personal computer (Intel i3 chipset, 4GB RAM).

The neural network is then followed by a hidden layer with 15 neurons defined by the Kolmogorov expression (Hetch-Nielsen, 1989):

N2n+1

Where: N-Number of neurons in the hidden layer; n-Number of inputs.

The output layer of a single neuron related to the back pressure of drying chamber, which leads us to the bed permeability.

We used the error backpropagation training algorithm, the learning function of gradient descent and the Round mean squared error (RMSE) as the performance function.

The transference functions were respectively hyperbolic tangent for the hidden layer and logistic for the output layer.

The sensitivity analysis was conducted to infer the relative importance of each variable. This contribution was measured by the RMSE degradation after the application of a 2%, 5% and 10% noise on each variable.

The neural network modeled for this application is summarized in Figure 1.

Figure 1 Artificial Neural Network, 7-15-1.  

3. Results and discussion

After the artificial neural network validation with industrial data, there was a convergence as shown in Table 2 with the mathematical model.

Table 2 correlation coefficients R2

Training Test
R2 0.870159 0.865757
RMSE 0.036486 0.0329154

Regression between the target and the Artificial Neural Network predicted outputs is plotted in Figure 2, pointing to a good convergence considering a factory scale experiment.

Figure 2 Regression for testing data.  

The sensitivity analyses of the input variables were made exhibiting indexes in Figure 3. It was observed that the variables of highest influence on the drying chamber pressure and consequently in pellet bed permeability was the diameter standard deviation, followed by pellet production and pellet sphericity.

Figure 3 Variable sensitivity.  

Particle size distribution, which has a direct relationship with the standard deviation of the pellet diameter, has a main role in bed permeability, since smaller pellets occupy interstitial sites among bigger pellets, decreasing bed voiding (Koekemoer and Luckos, 2015). In his way, a wider diameter distribution and/ or the presence of fines block the air flow through the bed (Yazdanpanah et al., 2010), causing a drying deficiency on the opposite side. A better control of diameter distribution can only be performed during the pelletizing step, showing the importance of control parameters during this phase.

According to the presented Ergun equation, bed height has a direct proportion on bed permeability, but controlling this factor in order to reduce back pressure will result in productivity reduction. In this case, despite a great influence in the final result, bed height should not be the parameter to be controlled.

Sphericity also affects permeability, since the pellet shape is decisive for bed formation and pellet packaging. The particle size and shape influence on bed formation is the object in several studies (Al-Raoush and Alsaleh, 2007; Li and Ma, 2011; Allen et al., 2013; Kruggel- Emden and Vollmari, 2016). It is possible to infer that low compression strength green pellets tend to deform easily in contact with other pellets and/or in contact with a traveling grate kiln, forming a single block that chokes the air flow through the bed. Studies performed to check shape influence on bed formations have so far considered rigid particles. Such characteristic may be caused by a poor control of raw material and excess of water in the pulp.

Permeability is also connected with other aspects, such as, particle roughness, tank to particle ratio, and gas flow speed (Trahana et al., 2014).

Mean size and pellet diameter ranging above 16mm and below 10mm show minor influence on the final result, proving that diameter distribution and particle shape modulations have a higher influence on bed permeability. Pellets distributed in a narrow diameter with spherical shape lead to a better bed formation.

The prediction model for backpressure in the drying chamber has its results plotted in Figure 4, presenting a good correlation considering the factory conditions where the data where obtained.

Figure 4 Pressure prediction, Target / Output.  

4. Conclusions

The model pointed out variables that significantly affect the bed permeability in the drying stage of iron ore pellet production. Diameter standard deviation, pellet production and sphericity were identified as the main effective input variables that affect bed permeability during the drying stage and consequently, final product quality.

The artificial neural network method showed a suitable option in mathematical modeling of a complex process involving a lot of variables under a real production site environment. The computational model was able to predict optimizations in the preparation, production and quality control of pellet production, despite all the interrelated factors of a production plant.

A higher diameter standard deviation and the presence of fines negatively affect the back pressure inside a travelling grate kiln (Luckos and Bunt, 2011), since smaller particles occupy interstitial sites among bigger pellets.

Sphericity has an important role in bed formation, but its aspects should be better understood, since iron ore green pellets are subject to plastic deformation during their generation process.

Acknowledgements

This work was supported by the Instituto Federal do Espírito Santo, Vale S/A, and CNPq.

References

AL-RAOUSH, R., ALSALEH, M. Simulation of random packing of polydisperse particles. Powder Technology, n. 176, p. 47–55, 2007. [ Links ]

ALLEN, K.G., VON BACKSTRÖM T.W., KRÖGER D.G. Packed bed pressure drop dependence on particle shape, size distribution, packing arrangement and roughness. Powder Technology, n. 246, p. 590–600, 2013. [ Links ]

DWARAPUDI, S., GUPTA, P. K., RAO, S. M. Prediction of Iron Ore Pellet Strength Using Artificial Neural Network Model. ISIJ International, v. 47, p. 67-72, 2007. [ Links ]

ERDIM, E., AKGIRAY O., DEMIR, I. A revisit of pressure drop-flow rate correlations for packed beds of spheres. Powder Technology, n. 283, p. 488–504, 2015. [ Links ]

HAYKIN, S. Neural networks and learning machines. (3. ed.). New Jersey: Pearson, 2009. [ Links ]

HECHT-NIELSEN, R. Neurocomputing. (3. ed.). Michigan: Addison-Wesley Publishing Company, 1989. 433 p. [ Links ]

HINKLEY, J., WATERS, A.G., O'DEA, D., LITSTER, J.D. Voidage of ferrous sinter beds: new measurement technique and dependence on feed characteristics. International Journal of Mineral Processing, n.41, p.53-69, 1994. [ Links ]

HOLLANDS, K. G. T., SULLIVAN, H. F. Pressure drops across rock bed thermal storage systems. Solar Energy, v. 33, n.2, p. 221 225, 1984. [ Links ]

KOEKEMOER, A., LUCKOS, A. Effect of material type and particle size distribution on pressure drop in packed beds of large particles: Extending the Ergun equation. Fuel, n.158, p.232–238, 2015. [ Links ]

KRUGGEL-EMDEN, H., VOLLMARI, K. Flow-regime transitions in fluidized beds of non-spherical particles. Particuology, 2016. (in press). [ Links ]

LI, L., MA, W. Experimental characterization of the effective particle diameter of a particulate bed packed with multi-diameter spheres. Nuclear Engineering and Design, n.241, p.1736–1745, 2011. [ Links ]

LUCKOS, A., BUNT, J.R. Pressure-drop predictions in a fixed-bed coal gasifier. Fuel, n.90, p.917–921, 2011. [ Links ]

LOPES, N., RIBEIRO, B. An Efficient Gradient-Based Learning AlgorithmApplied to Neural Networks with SelectiveApplied to Neural Networks with Selective. Neural, Parallel and Scientific Computations, n. 11, p. 253-272, 2003. [ Links ]

MATOS, A. P. Influência da temperatura, pressão, produção e granulometria no processo de secagem das pelotas cruas. 2007. (Dissertação de mestrado). [ Links ]

MEYER, K. Pelletizing of iron ores. Heidelberg: Springer-Verlag Berlin, Dusseldorf: Verlag Stahleisen mbH, s.n., 1980. [ Links ]

THELLA, J. S., VENUGOPAL, R. Modeling of iron ore pelletization using 3 ** (k–p) factorial design of experiments and polynomial surface regression methodology. Powder Technology, n. 211, p. 54-59, 2011. [ Links ]

TIMOFEEVA, A. S., NIKITCHENKO, T. V., ROGOVA, A. L. Optimizing the granulometric composition of pellets obtained on a disc pelletizer. Metallurgist, Moscow, n. 5, p. 48-51, 2013. [ Links ]

TRAHANA, J., GRAZIANIB, A., GOSWAMIA, D. Y., STEFANAKOSA, E., JOTSHIA, C., GOELC, N. Evaluation of pressure drop and particle sphericity for an air-rock bed thermal energy storage system. Energy Procedia, n.57, p. 633 – 642, 2014. [ Links ]

YAZDANPANAH, F., SOKHANSANJ, S., LAU, A.K., LIM, C.J., BI, X., MELIN, S.,AFZAL, M. Permeability of wood pellets in the presence of fines. Bioresource Technology. n.101, p.5565–5570, 2010. [ Links ]

Received: March 21, 2016; Accepted: September 26, 2016

Creative Commons License This is an Open Access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.