Académique Documents
Professionnel Documents
Culture Documents
Abstract—This paper presents the development of an artificial forming process is expected to improve greatly. This process
neural network model for the prediction of heating-line is an important production process that can be widely used to
positions in induction heating process. This model helps produce various curved thick plate for ship industry.
shipyard manufacturers determine the positions of induction However, we do not yet have enough strong tools not only to
heating lines and their heating parameters to form a desired determine the heating parameters including heating positions
shape of plate. The vertical displacements of plate have been to form a desired shape but to determine the deflections of
considered as the input parameters and the selected induction plate with the heating parameters. The previous studies did
heating lines as output parameters to develop the model. The not give any method for determining net of induction heating
training patterns of neural network are obtained using an
flux. Furthermore, travel of the inductor over a workpiece
analytical solution that predicts plate deformations in
was not modeled in any previous studies for simulation of an
induction heating process. The developed neural network
model is tested to show its feasibility to determine the heating actual induction process. Therefore, the induction heating
positions on the surface of a flat steel plate in the line heating process should be precisely modeled to evaluate its
process for forming a desired shape. feasibility in plate forming. In addition to this, most
importantly, to apply the process to the steel-plate forming,
Keywords- shipyard; Neural Network; induction heating; line the relation between positions of heating line and desired
heating; forming plate shape of the plate should be necessarily known in advance.
In this study, with the simplified formula, an Artificial
I. INTRODUCTION Neural network (ANN) model is proposed to generate the
heating-line positions and the heating parameters from the
The steel forming process in shipyard is considered as objective curved surface. ANNs are powerful tools that have
an important stage with respect to productivity and precision recently emerged for modeling complex real-world
of curved plates. Two types of heat sources can be used in problems. ANNs are composed of densely interconnected
the line heating process: oxyacetylene torch and electro- adaptive simple processing elements, called artificial
magnetic induction. The thermo-mechanical forming using neurons or units that are capable of executing massively
an oxyacetylene torch is more versatile and less expensive, parallel computations for data processing and knowledge
but it depends exactly on the skill of experienced workers to representation. The advantage of the artificial neural
control the deformation to produce a curvature within an networks stems from the remarkable information processing
allowable range. To control the deformation precisely characteristics of the biological system such as nonlinearity,
during the process, the heat flux from the source should be high parallelism, fault and failure tolerance and learning.
easy to control. However, the heat flux from the The main purpose of artificial neural network which is
oxyacetylene mixed gas is not easy to control because of the based on computing is to develop mathematical algorithms
heat generation characteristics resulting from the reaction of that will enable artificial neural networks to learn by
the gases. Meanwhile, because the dimension of hull needs mimicking information processing and knowledge
to be controlled more precisely for high-performance ships, acquisition in the human brain.
the oxyacetylene process may be limited in various
applications. An alternative heat source of electro-magnetic
induction has been suggested as substitute for the gas flame
heating for such applications. Compared with the heat
source of an oxyacetylene torch, the induction process has
the following advantages as the power and its distribution
are easier to control and reproduce. Besides the induction
system can be integrated with a robotic system for
automation.
Induction heating process is known to produce
controllable heat on a conductive workpiece. When the
induction heating process is applied in association with Figure 1. Schematic model of induction heating process.
automatic inductor-handling equipment and a heating line
generation algorithm, the productivity of the curved plate
135
TABLE I. OUTPUT VARIABLES FOR AN OUTPUT LINE paths are indexed shown in Fig. 3, and they are able to be
Parameter Value Meaning horizontal lines, vertical lines, or diagonal lines along which
Parameter 1 0 Induction heat the inductor moves. Values of output layer are set to be
1 No induction heat binary. Each induction heating line is predicted with 4
Parameter 2 0 Back surface parameters as shown in Table I. Values of output layer can
1 Front surface be binary, 1 or 0. Each induction heating line is predicted
0 0 1st velocity
Parameter 3 +
with 4 parameters as:
0 1 2nd velocity
Parameter 4 1 0 3rd velocity
- Parameter 1: Active Path, if value is 1, there is the
1 1 4th velocity heat at induction heating line and one is 0 then no heat at
line.
TABLE II. ENCODED VELOCITY - Parameter 2: Surface parameter, if value is 1 then line
High bit Low bit Velocity(m/s)
heating in front surface and 0 is heated back surface.
1 1 20 - Parameters 3 and 4: Velocity, 2 remaining values used
1 0 15 for a path showing the 4 inductor velocities: they encode 4
0 1 10 values of velocity (Example: encoding 00, 01, 10, 11 are
0 0 5 equivalent 5, 10, 15, 20 (m/s)).
In Fig. 3, there are (3n-5) induction heating paths,
III. MODELING FOR PREDICTING INDUCTION which can be generated on plate. Therefore, there are (3n-5)
HEATING PATHS USING NEURAL NETWORK paths using for predicting in this application. Each path
ANNs are computational models, which replicate the includes 4 parameters as above shown in Table I and there
function of a biological network, composed of neurons and are 4(3n-5) components in output layer. Therefore, there are
are used to solve complex functions in various applications. (3n-5) categories of the output vector, each representing
ANNs consist of simple synchronous processing elements several groups of induction heating lines. Each path number
that are inspired by the biological nerve systems. The matches position of induction heating line. When first
general network in this study has 3 or 4 layers, which are parameter of group i (ith induction heating path) is 0, it is
input, output, and 1 or 2 hidden layers. The basic units in meaning no induction heating path at the line i and therefore
ANN are neurons, which are connected to each other by velocity, and heat surface are also 0.
links known as synapses. Associated with each synapse, Because the output value for each simulation’s case must
there exists a weight factor. One of the most popular be either 1 or 0, therefore calculated value at output unit after
training algorithms is the back-propagation algorithm. The through network is compared with threshold like:
hidden layers of neurons receive the inputs from the ⎧ ⎛ h2 ⎞
previous layer and convert them to an activation value that ⎪ 1 if f ⎜ ∑ zi wik ⎟ ≥ 0.5
can be passed on as input to the neurons in the next layer. ⎪ ⎝ i =0 ⎠ (2)
The input-layer neurons do not perform any computation, yj = ⎨
⎪0 if f ⎛ z w ⎞ < 0.5
h2
they merely copy the input values and associate them with
⎪ ⎜ ∑ i ik ⎟
weights, feeding the neurons in the hidden layer. ⎩ ⎝ i =0 ⎠
In the induction heating process, the deformed shape of In the following, modeling for predicting induction
plate is changed due to the number and position of induction heating paths with the back-propagation network (BPN) will
heating line. In the proposed model, ANN is used to solve be discussed. The units of input layer are used to receive the
the problem selecting the positions of induction heating path (n2) displacement values. The units of output layer are used
with heated surfaces, and velocities. The network is trained to send out the active induction heat enclosed with heating
for positions of induction heating line and deformation of surface, and velocity. By trial, the mean squared error
plate after heated. Initial shape is a square plate, which is (MSE) values are calculated to get minimum error as well as
divided into n2 points to get displacements (deflections) stability.
with equal distance from a point to other on the planar area One of the most important works in ANN studies is to
as shown in Fig.2(a). The point i’s vertical displacement of determine the optimal network architecture which is related
gridded plate along z-axis as shown in Fig.2(b) is hi. to the number of hidden layers and neurons in it. ANN
Therefore, each input vector has n2 units representing structure has significant effects on the predicted results.
deformed plate’s shape expressed as a pattern on a (nxn) However, the optimal number of hidden layers and the
grid of displacement’s point. optimal number of neurons in each layer are dependent on
The plate’s angular distortion in induction heating specific case and there is no straightforward method for
process is function of some parameter as heating speed, determination of them. Generally, the trial and error
plate thickness as (1). As values of output units, induction approach is used. Number of hidden layers with specific
heating paths are needed to generate the deformed plate. models depends on number of training patterns and
The output layer is recognized by parameters as number of accuracy of classification expected. And theoretical results
induction paths, heated surfaces, and velocities. A heating show that one hidden layer is sufficient for back-
path includes heated line (heat or no heat) with velocity of propagation algorithm to map any input to each output to an
inductors, and heated surface (front or back) of path. The arbitrary degree of accuracy.
136
Figure 6. Induction heating lines on front surface (left) and back surface
(right) in experiment.
(a)
1000
Deflection (mm)
137
highest accuracy for predicting induction heating was attributed to the following reasons: first, due to number of
obtained at a network, which has 1 hidden layer with 160 training data and random distributions are not good for
neurons. It was found that increase of the number of neurons predicting parameters in this case. To improve the problem,
in the hidden layer does not ensure the decrease of the an ANN model is trained with lager number of data and
average training process’s error or increase accuracy. Based patterns are chosen randomly with equilateral distributions as
on this analysis, the optimal architecture of the ANN was well as combine of more induction heating line than 10.
constructed as 121-160-112 presenting the number of Secondly, 4 chosen velocities have large different levels,
inputs, neurons in hidden layers, and output, respectively. therefore, they can affect on errors of deformed plate. To
Training data set with 1200 patterns as database was reduce the errors, number of bits to encode velocity used
selected random based on relationship between the should be more than 2 bits or differences between 2
deformed shapes of plate and positions of induction heating velocities should be smaller. The improvements will bring
line as well as their velocities and heated surface. The better results for predictions of induction heating paths.
vertical displacements at the points on the plate were Nevertheless, the present result shows that the developed
calculated as input values according to the simple formula model can help to predict the induction heating paths from
(1). analytical displacements, and it is indeed a powerful tool to
In this case, training a neural network involves 10,000 established relationships between vertical displacements and
iterations (also known as epochs). Database randomized is induction heating paths.
used to determine the weight and biased values. The
learning rate was set at 0.2 and momentum coefficient at V. CONCLUSIONS
0.8. Training was performed until the iteration number of The main conclusions from this study are summarized
10,000 times and error 0.001 was reached. At the beginning as follows.
of training, ANN initializes network weights with a set of - A simplified formula for easily simulating plate
small random values according to Nguyen-Widrow deformations in induction heating process has been
initialization [4]. During each epoch, the network received a developed with a series of FEM analyses of the thermo-
sequence training pairs. Then, the system calculated a sum mechanical behavior. Fairly good agreement between the
squared error between the desired outputs and the network’s simulating and experimental results has been achieved.
actual outputs. It used the gradient of the sum squared error - An ANN model capable of predicting line-heating
(with respect to weights) to adapt the network weights so paths with the necessary heating parameters is developed
that the error measure is smaller in next epochs. The with the back-propagation algorithm. The feasibility study
accuracy of the developed model depends on these weights. proved the model as a predictive tool for determining heating
Training terminates when the sum squared error is below a paths from object shapes can produce the actual shapes of
specified tolerance limit, optimum weights are reached, the plates in induction heating process.
weights and biased values encode the network’s state of - The developed model can effectively reduce the time
knowledge. required to solve the problem of matching deformation of
After training, a BPN is simulated by using only the feed- plate with induction heating line positions and the problem
forward phase of the trained algorithm. As a specific case, of predicting induction heating paths.
from desired surface, displacements of points were given as - The algorithm of the ANN obtained from this study,
shown in Fig.5. The values of them were saved at input which is the heating line generation algorithm on how and
vector for simulating ANN. Next, the induction heating where to heat a workpiece, can be suitably and reliably
experiments on the AH32 plate with the predicted induction applied to the induction heating process with larger plate and
heating paths and parameters are conducted to evaluate the more paths.
proposed ANN model. This induction heating process is
subdivided into 2 sub-processes as front surface and back ACKNOWLEDGMENT
surface. There are 2 parallel induction lines in front surface This study was financially supported by Chonnam National
of plate and 3 parallel lines in back surface as shown Fig.6. University, 2009.
The initial shape of plate is planar, after the plate was heated
along 5 predicted lines. The vertical displacement values of REFERENCES
plate at points were measured by a Coordinate Measuring [1] V. Rudnev, D. Loveless, R. Cook and M. Black, “Handbook of
Machine (CMM). The desired(target) and experimental Induction Heating,” Marcel Deckker 2003.
results for vertical displacements when induction heating [2] Y. Favennec, V. Labbe and F. Bay, “Induction heating processes
process was performed on flat plate 20 mm thickness are optimization a general optimal control approach,” J. of Computational
shown in Fig.7, it shows vertical deflection of plate y = -500, Physics; Vol. 187, 2003, pp.68-94.
-300, -100, and 300 (mm) for simulating the specific case. In [3] A. Masserey, J. Rappaz, R. Rozsnyo and R Touzani, “Power
formulation for the optimal control of an industrial induction heating
the graphs, the experimental results are almost near the process thixoforming,” Inter. J. of Applied Electromagnetics and
desired results. From the graphs in Fig. 19, it is obvious that, Mechanics, Vol. 19, 2004; pp. 51-56.
although there are some errors between desired and [4] D. Nguyen and B. Widrow, “Improving the learning speed of 2-layer
experimental displacement values, the shape of predicted Neural Networks by choosing initial values of the adaptive weights,”
plate is very similar to the desired one. The differences Proc. Inter. Joint Conf. on Neural Networks, San Diego, 1990.
between the desired and experimental values may be
138