Application of the neural networks for modeling the molten steel level variation in the continuous casting process
K. Gherfi a, S. Bouhouche a
a Unité de recherché appliquée en Sidérurgie et Métallurgie URASM-CSC.
B.P 196 Annaba, 23000, Algerie kaddour.gherfi@gmail.com Keywords: modeling, continuous casting, neural networks
Extended Abstract
In this paper our principal study is the development of a model of synthesis presents the variation of the molten steel level in the ingot mould of continuous casting by using the extraction speed in the input of the model. The approach used is based on the neural networks technique which allows modeled the process of level variation, and also makes it possible to know the reliability of the neural networks the modeling the industrial processes.
I Introduction
In the steel industry sector, the casting continuous takes a very significant role in the production.
Among the traditional methods of this process, the cast in the ingots. It is less expensive, its increased productivity.
This process appeared in the imagination of considerable researchers in the middle of the 19 century. It had been necessary wait almost a century to see a machine cast a ferrous metal in the industry [1].
Among the principal factors which contains in the continuous casting process, there are two things, the molten steel level in the ingot and the extraction speed, the variation of these two factors influence directly on the quality of metal. For that we interested here on the modeling of the molten steel variation by using the extraction speed in the input of the model
II the continuous casting
The machine of continuous casting is represented in Fig. 1.
The aim of such machine is to from continuously, in the open ingot at its two ends, a carapace of solid of metal resistant to collect the molten metal, then to advance this carapace by completing the solidification by water sprinkling.
Figure 1 : Machine of continuous [2].
Physically, the machine of continuous casting is formed by an ingot (see Fig. 1) and a hollow bottomless delimited by the rollers. The distributor and the ladle make it possible to supply the casting machine of liquid steel with a controlled flow.
The majorities of the rollers are only used to maintain a certain pressure on the solid crust and thus supported completely the slab. However, there are rollers which are provided with engines to advance the slab. In addition the engineers have also the possibility of carrying out a soft reduction (i.e. to compress the slab between two rollers in opposite) near the area of the end of the total solidification product. Practically, the reduction allows to modify the thermomechanical state at the end of solidification and limited certain types of defects [1].
III The neural networks
The first neural architecture is the perceptron. It appears in 1958, it is the result of Rosenblatt works [3]. Artificial neural network (ANN) is composed typically by a layer in the input, one or several layers intermediate (hidden) and a layer in the output. In the ANN, each node or neuron is connected to other neurons of a preceding layer by adaptable synaptic weights.
Fig. 2 shows the general diagram of a neuron. Each node i calculates the sum of its inputs (Pi…, PN), balanced by the corresponding synaptic weights (wi1…, wiN) and added the result to the bias b, this value represents the internal state of the neuron (ui). This result is then transmitted by an activation function f (Fig. 3 represented the most activation functions used in neural networks). The output yi is the activation of the node [4].
∑
=
+
=
N
n n in
i
w P b
U
1
,
y
i= f ( u
i)
(1)The expression using to calculate the new values of synaptic weights connecting the neurons is given by (2) [5].
j i ij
ij
k W k D P
W ( + 1 ) = ( ) + λ
(2) With:λ
: Step of training.P
j: The input of neuron j.W
ij: Weight associated to the connection of neuron I with the neuron j.D
i: Derivative of the error of neuron i.The ANN allows to approach the nonlinear relations with significant degrees of complexity. The input cells are intended to collect the information which is transformed by the hidden cells to the output cells. These networks have one or more hidden layers (Fig. 4). In this type of networks, a sigmoid activation function is generally used [6].
) exp(
1 ) 1
( x x
g + −
=
(3)Figure 2 : The general diagram of a neuron.
Inputs A neuron with R inputs
Figure 3 : Activation functions: (a) sigmoid function, (b) linear function, (c) function with
threshold, (d) Gaussian function [4].
In this study, the behavior of the molten steel level variation is modeled by using the neural networks. For the training of our network with the input-outputs data, the propagation algorithm is used.
After training and obtaining the convergence, the obtained model will be used for the prediction by a new data base.
The input-outputs parameters of our neural network are the extraction speed in the input and the molten level variation in the output.
IV Data acquisitions
We acquired the data from the continuous casting of ArcelorMittal of Algeria by Memograph M, RSG40 constructed by Endress Hauser, connected with a PC. The ReadWin® 2000 software is used. Fig. 9 and 10 present the measurements.
The measurements are taken in the case of absence of defects.
V Results and discussion
Our network is constituted by one (01) input [Speed] and one (01) output [Level]. In our case we used the MLP (multi layer perceptron) networks in which we have:
-One (01) neuron in input,
-Ten (10) neurons in hidden layer, their activation function is sigmoid type, -One (01) neuron at output, its activation function, is linear type.
A. Training phase
The training phase is executed by using the data without relative defect of three hundred (300) points of measurement (samples) where the network is tested by the same data.
The training is considered as completed after 200 times (an optimal number obtained by simulation).
Figure 4 : Multi-layer networks MLP.
Hidden layers Output layers Input layers
Molten steel level Extraction
speed
Neural network model (NN)
Figure 5 : Neural network structure.
0 50 100 150 200 250 300
1.8 2 2.2 2.4 2.6 2.8 3
Extraction speed (m/min)
Sampling number
Figure 6 : Variation of speed extraction as a function of time.
0 50 100 150 200 250 300
500 550 600 650 700 750
Molten steel level (mm)
Sampling number
Figure 7 : Variation of molten steel level as a function of time.
We observe that the molten steel level variation obtained using the NN (neural network) model follow the same behavior as the real values of this variation.
We show that the neural model obtained satisfies the objective initially set for training.
In the continuous casting process the decrease of the signal represent the molten steel level variation to the zero, means that there is a defect in the process.
In this part, we try to model a case where a defect is present.
B. validation Phase
The same parameters used in the training phase, are used here, the prediction was executed by a new data base contains a defect.
We visualize the error between the real values and the values calculated by the neural network.
From Fig. 10 we observe that the neural network gives a good result for the prediction. In fact even if the neural network model obtained in the training phase without defects, in the validation phase, this model detects the defect in the case of its presence.
Error of modeling
Sampling number
0 50 100 150 200 250 300
-60 -40 -20 0 20 40 60
Figure 9 : Variation of the training error.
0 50 100 150 200 250 300
500 550 600 650 700 750
Molten steel level (mm)
Sampling number Real values
NN model
Figure 8 : Prediction of molten steel level variation using neural network.
0 50 100 150 200 250 300
0 100 200 300 400 500 600 700
Molten steel level (mm)
Sampling number Real values
NN model
Figure 10 : Prediction of molten steel level variation using neural network.
0 50 100 150 200 250 300
-80 -60 -40 -20 0 20 40 60 80
Error of modeling
Sampling number
Figure 11 : Variation of the validation error.
VI Conclusion
In this paper, we were interested in the molten steel level variation in the ingot for continues casting process. In particularly, we used the neural network model for modeling of this variation. For this, the extraction speed is used in the input of the model.
The neuronal approach gives a good modeling of the molten steel level variation in the two cases presence or absence of the defect.
References
[1] Alban HEINRICH. Modélisation thermomécanique de la coulée continue d’aciers en deux dimensions. Thèse de doctorat, École Nationale Supérieure des Mines de Paris, 2003.
[2] K. Jabri, D. Dumur, E. Godoy, A. Mouchette, B. Bele, "Modified Smith Predictor Scheme for Periodic Disturbances Reduction in the Continuous Casting Process" IEEE Xplore, pp. 1091-1096, 7th ASCC, Hong Kong, China, Aug. 27-29, 2009.
[3]
Perez-Uribe, A. (1998). Artificial Neural Networks: Algorithms, Hardware. Implementation. dans D.
Mange, M. Tomassini, Bio-Inspired Computing Machines. Toward novel comutational architectures, Presses Polytechniques et Universitaires Romandes.
[4]
Lyes BADJI, "Contribution au Développement d’un Système d’Aide à la Surveillance de Patients Hospitalisés en Unités de Soins Intensifs", Thèse de doctorat, Institut National Polytechnique de Grenoble, 2006.
[5]
Y. HAMMOU, M. ZERIKAT, A. BELAIDI, “Application des réseaux de neurones pour l’identification d’un moteur à courant continu en vue de son contrôle “,COMAEI’96, Tlemcen, 1996
.[6]