You are on page 1of 8

Indian Journal of Engineering & Materials Sciences

Vol. 13, February 2006, pp. 37-44

Optimization of friction welding parameters using simulated annealing


P Sathiyaa, S Aravindanb, A Noorul Haqb & K Panneerselvamb
a
Department of Mechanical Engineering, J.J. College of Engineering & Technology, Tiruchirappalli 620 009, India
b
Department of Production Engineering, National Institute of Technology, Tiruchirappalli 620 015, India
Received 6 August 2004; accepted 1 August 2005

Friction welding is a solid state bonding technique which utilizes the heat generated by rubbing of two faying surfaces
for the coalescence of material. Friction welding process is normally considered as a multi-input and multi-output process.
Optimization of friction welding parameters is required to get good quality joints. Present study establishes the relationship
between the input variables and output variables. Heating pressure (HP), heating time (HT), upsetting time (UT) and
upsetting pressure (UP) are the input variables. The flash features such as flash thickness (FT), flash height (FH) and flash
width (FW) are considered as output variables. The relationship between these input and output variables is established
through artificial neural network (ANN). ANN is suitably integrated with the optimization algorithm in order to minimize
the metal loss without sacrificing the tensile strength of the joints. The variation between theoretical and experimental values
of flash features is also analyzed.
IPC Code: B23K20/12

Friction welding is a solid state welding process. It


makes use of frictional heat generated at the rubbing
surfaces to raise the temperature of the interface
higher enough to cause the two surfaces to be forged
together under high pressure. Continuous drive
friction welding in which the energy required to make
the weld is supplied by the welding machine through
a direct motor connection for a preset of the welding
cycle (shown in Fig. 1). In continuous drive friction
welding, one of the work pieces is attached to a motor
driven unit while the other is restrained from rotation
(Fig. 1a). The motor driven work piece is rotated at a
predetermined constant speed. The work pieces to be
welded are forced together (Fig. 1b) and then a
friction force is applied (Fig. 1c). Heat is generated,
as the faying surfaces (weld interface) rub together,
this continues for a predetermined time or until a
Fig. 1—Continuous drive friction welding process
preset amount of axial shortening (upset) takes place.
The rotational driving force is discontinued and the
rotating work piece is stopped by the application of a
braking force. The friction force is maintained or
increased, for a predetermined time after rotation
ceases (Fig. 1d). Figure 2 explains the variation of
welding speed, friction force and forge force with
time. The stages of welding such as starting, axial
shortening and weld completion are also explained in
the same figure.
The input variables that control the joints are
heating pressure (HP), heating time (HT), upsetting
_____________
*For correspondence (E-mail: paulsathiya@yahoo.co.in Fig.2—Parameters on continuous drive friction welding
38 INDIAN J. ENG. MATER. SCI., FEBRUARY 2006

pressure (UP) and upsetting time (UT). The output (FW), flash height (FH) and flash thickness (FT) in
variables are flash width (FW), flash height (FH) and friction welding, we apply the non-traditional
flash thickness (FT). To process good quality joints, optimization algorithm called simulated annealing. A
these parameters have to be optimized. Optimization generalized objective function is established.
of friction welding parameters will be of time- Artificial neural network (ANN) technique is suitably
consuming if the conventional technique of integrated with simulated annealing. Simulation
optimization is used, by concentrating on a single results confirm the feasibility of this approach, and
parameter whereas keeping the others as constant. show a good agreement with experimental results for
The main consideration in this study is to minimize a wide range of joining conditions.
the flash width, flash height and flash thickness,
without sacrificing the strength of joints. There have Experimental Procedure
been many studies on screening experiments, A continuous drive friction-welding machine with a
modelling and optimization for welding processes. maximum 150 kN load was used for welding. The
Chang1 reported the fundamental bonding friction and forge pressures in the range of 15-25 bar
mechanisms of ultrasonic welding. An optimum weld and 35-45 bar respectively. The spindle rotating speed
condition for the electric power input was formulated was kept constant at 1125 rpm and the welding was
using elastic and plastic analysis. A semi-analytic performed under the specified friction upset distance.
mathematical expression for the optimum electric Austenitic stainless steel (AISI 304) specimens of size
power input was obtained. Graville2 studied on 16 mm diameter and length 130 mm were used as
establishing the minimum size single pass fillet welds parent materials in this study. The chemical
that can be deposited without a preheat and without composition of the specimen material is presented in
HAZ cracking using the submerged arc welding Table 1. Similar austenitic stainless steel specimens
process. The fillet weld size was experimentally were joined by friction welding process without any
optimized based on the hardness criterion of heat- preheat. The friction welding parameters for this
affected zone (HAZ). Zhang3 reported polar study is presented in Table 2. Theoretical
coordinate model to characterize the weld pool optimization was carried out in order to minimize the
geometry. A neural network was therefore proposed flash width, flash height and flash thickness of the
to identify the parameters in real time. By using joint by simulated annealing. The process was
pulsed laser illumination, clear images of the weld considered here as multi-input and multi-output
pool could be captured. The developed image- system. Flash parameters play an important role in
processing algorithm extracts the boundary of the determining properties of the weld. Theoretical and
weld pool in the real time. Elena Koleva4 carried out experimental variations in the flash width, flash
statistical analysis of electron beam welding of height and flash thickness of the joint were also
stainless steel samples using the multi response predicted. Mathematical equation was formulated to
statistical techniques. A model was created which represent the objective function and backpropagation
included the values of the distance between the neural network was designed and trained to have the
electron gun and both the focusing plane of the beam relationship between the input parameters (HP, HT,
and the sample surface as parameters. Recently, UP and UT) and output parameters (FW, FH and FT).
evolutionary computational techniques are used for The relationship obtained between the input and
optimizing the welding parameters. Tarng5 has
reported on the usage of neural network to construct Table 2— Range of input variables
the relationships between welding process parameters
Sl. No. Input variable Range
and weld pool geometry in tungsten inert gas (TIG)
welding. However, there have been few techniques to 1 Heating pressure 15-25 bar
move the experimental region to near optimal welding 2 Heating time 3-10 s
conditions. In order to determine the welding process 3 Upsetting pressure 35-45 bar
parameters that produce the optimized flash width 4 Upsetting time 3-7 s

Table 1— Base material chemical composition

Element C Si Mn P Cr Ni Co Mo As Pb Ti Fe
% 0.0468 0.3446 1.313 0.0182 17.87 8.289 0.0785 0.010 0.0152 0.0007 0.0355 Balance
SATHIYA et al.: OPTIMIZATION OF FRICTION WELDING PARAMETERS 39

output parameters by artificial neural networks Bounds on heating time


(ANN) was optimized by using simulated annealing.
HTL ≤ HT ≤ HTU …(3)
General Model
Process modeling and optimization are very where HTL and HTU are the lower and upper bounds
important issues in welding engineering. Because of of heating time respectively.
the complexities involved in welding process,
optimization and optimal control of the process are Bounds on upsetting pressure
difficult to be performed. Therefore, ANN is used, to
map the input/output relationships of the process. The UPL ≤ UP ≤ UPU … (4)
configuration of the backpropagation network for
friction welding is shown in Fig. 3. An objective
where UPL and UPU are the lower and upper bounds
function is formulated in terms of predictor variables.
of upsetting pressure respectively.
The objective function f can then be defined as
Bounds on upsetting time
f =W1 (FW) + W2 (FH) + W3 (FT) … (1)

where, W1, W2 and W3 are the weights for the UTL ≤ UT ≤ UTU ... (5)
normalized flash width [FW], normalized flash height
[FH] and normalized flash thickness [FT] of the weld, where UTL and UTU are the lower and upper bounds
respectively. Experiments were conducted at random of upsetting time respectively.
choice and flash width (FW), flash height (FH) and Bounds on tensile strength
flash thickness (FT) were then measured for each set
of data. From the experimental data, the ANN is
TS L ≤ TS ≤ TS U ... (6)
trained.

Welding constraints where TSL = 520 MPa and TSU = 600 MPa are the
The practical constraints imposed during the lower and upper bounds of tensile strength
welding operations by taking into consideration of the respectively.
range available for the machine parameters. Also an
added constraint is imposed, in order to retain the Simulated annealing algorithm
strength of joints, while minimizing the metal loss. Traditionally, the annealing process used in
metalworking involves heating the metal to a high
Parameter bounds temperature and then letting it gradually cools down
Bounds on heating pressure to reach a minimum stable energy state. Metropolis et
al. 6 proposed a criterion to simulate the cooling of a
HPL ≤ HP ≤ HPU …(2) solid for reaching a new energy state. Based on the
Metropolis criterion, a search algorithm called
'simulated annealing' was developed7,8. Simulated
where HPL and HPU are the lower and upper bounds
annealing is a Monte Carlo approach, used for
of heating pressure respectively.
minimizing multivariate functions. The term
simulated annealing derives from the roughly
analogous physical process of heating and cooling a
substance to obtain a strong crystalline structure. In
simulation, a minima of the cost function corresponds
to this ground state of the substance. The simulated
annealing process lowers the temperature by slow
stage until the system “freezes” and no further
changes occur. At each temperature the simulation
must proceed long enough for the system to reach
Fig. 3—Configuration of the back propagation network for steady state or thermal equilibrium. It has been shown
friction welding that the simulated annealing algorithm possesses
40 INDIAN J. ENG. MATER. SCI., FEBRUARY 2006

several advantages in comparison with a traditional (0,1). If the r ≤ exp ( ΔE / Ts ) set Ts = Ts * Cr; Else go
search algorithm. First, the simulated annealing to Step 2.
algorithm does not need to calculate the gradient
descent that is required for most traditional search
where ΔE -difference between two consecutive
algorithms. This means that the simulated annealing
fitness value
algorithm can be applied to all kinds of objective and
Step 4: If the temperature is small, terminate. Else
constraint functions. Next the simulated annealing
go to Step 2.
algorithm with probabilistic hill-climbing
Figure 4 shows the flow chart for the simulated
characteristics can find the global minimum more
annealing. The process parameters are chosen in the
efficiently instead of becoming trapped in a local
same domain as shown in Table 2. The parameters
minimum where the objective function has
used in the simulated annealing algorithm are given
surrounding barriers. Furthermore, the simulated
as: the initial temperature Ts = 100°C, the final
annealing search is independent of initial conditions.
temperature Te = 0°C, the decaying ratio Cr = 0.8 and
As a result, the simulated annealing algorithm has
the weights W1 = W2 = W3 = 0.33.
emerged as a general optimization tool and has been
successfully applied in many manufacturing tasks9-11.
Results and Discussion
The SA procedure simulates this process of annealing
Typical macrograph of the friction welded sample
to achieve the minimum function value in a
is shown in Fig. 5. The metal loss in friction welding
minimization problem. The algorithm begins with an
is in the form of flash as observed in the macrograph.
initial point x1 ( HP1 , HT1 , UP1 and UT1 ) and a high Minimization of the metal loss without sacrificing the
temperature T. A second point
x2 ( HP2 , HT2 , UP2 , and UT2 ) is created using Gausian
distribution and the difference in the function values
(∇E) at these points is calculated. If the second point
has a smaller value, the point is accepted. The point is
accepted with a probability exp (-∇E/T). This
completes one iteration of the simulated annealing
procedure. The algorithm is terminated when a
sufficiently small temperature is obtained or a small
enough change in function value is obtained.

Simulated annealing steps


Step 1: Choose an initial point x1 , Set Ts a
sufficiently high value. Cooling rate Cr and set, Te = 0
Step 2: Calculate a neighbouring point using
Gausian distribution

⎡ n n⎤
x2 = x1 + σ ⎢ ∑ ri − ⎥
⎣ i=1 2⎦

Maximum value of parameter


− Mininum value of parameter
σ = variance =
6
where n is Number of random numbers, ri is random
numbers
( ) ( )
Step 3: If ΔE = E x ( t+1) − E x ( t ) > 0 , i.e., if the
difference in the fitness value is positive then, set Ts =
Ts * Cr; Else create an random number (r) in the range Fig. 4—Flow chart for the simulated annealing algorithm
SATHIYA et al.: OPTIMIZATION OF FRICTION WELDING PARAMETERS 41

strength of the joints is an important factor. It is also carried out. The variation of flash features with
possible to validate the joints by assessing the flash the process parameters such as heating pressure,
parameters such as flash thickness, flash width and heating time, upsetting pressure and upsetting time
flash height. The features of flash are explained in are presented in Tables 4-7 respectively. The
Fig. 6. Friction welding is carried out in a random theoretical and experimental results and the error
way; the process parameters and associated between them are also presented in the tables. The
experimental results are presented in Table 3. variation of flash features with the variation of
Artificial neural network is trained. Using the trained process parameters can be understood, from
ANN, the theoretical prediction of flash features was Fig. 7(a-c). It is observed that the increase in heating
pressure has an effect in the flash features. Heating
pressure for the given heating time, conditions the
interface. If the heating pressure increases,
thesoftened/conditioned material will be expelled.
Theoretically predicted values lie closer to the
experimentally observed flash features. The variation
of flash features with heating time is presented in
Fig. 8(a-c). With the increase in heating pressure, the
flash width, height and thickness are increased. The
effect of increase in heating time increases the metal
loss by way of increasing the dimensions of the
formed flash. The duration of heating is selected so as
to ensure the cleaned contact of faying surfaces by
friction. The temperature at interface is increased to
achieve the required softening for joints. When
Fig. 5—Friction welded sample heating time is too short, the heating effect become
irregular, and unbonded region increases. Increased
heating time results in increased metal loss. An
optimum level of heating time should be then selected
to minimize the metal loss.
Upsetting pressure in conjunction with surface
speed, determine the thermal conditions established in
the weld region. Also it dictates the amount of metal
loss extruded to form the collar. Increase in upsetting
pressure, increases the metal loss as shown in
Fig. 9(a-c). Increase in upsetting time also increases
Fig. 6—Features of flash the amount of flash as observed in Fig. 10(a-c). It can
Table 3—Experimental results for friction welding of AISI 304 stainless steel

Sl. No Heating Heating Upsetting Upsetting Flash Flash height Flash


pressure time pressure time width (mm) thickness
(bar) (s) (bar) (s) (mm) (mm)
1 15 3 40 5 9.56 7.58 4.6
2 20 5 40 7 11.38 8.2 4.87
3 20 8 45 3 10.7 8.15 4.33
4 15 5 35 3 9.86 7.98 4.21
5 16 5 36 3 12.62 8.46 5.81
6 17 6 37 4 12.87 8.92 5.76
7 18 6 38 4 12.86 8.83 6.01
8 19 7 39 5 13.15 8.34 5.98
9 20 7 40 5 13.87 9.01 6.27
10 21 8 41 6 13.92 9.21 6.42
42 INDIAN J. ENG. MATER. SCI., FEBRUARY 2006

be understood that theoretically predicted flash By carrying out experimental trials on this input
features by ANN are almost closer to experimental parameter, minimized metal loss can be obtained.
ones. Good agreement between the predicted and
measured values of flash width, flash height and flash Conclusions
thickness of the weld is observed from the presented Friction welding of similar AISI 304 stainless steel
tables. Hence the formulated ANN is resembling the is successfully performed. By way of conducting
real process. This model is integrated with simulated experiments on random selection of input parameters,
annealing algorithm in order to optimize the flash the flash features are measured. From the
features. The results of the optimization carried out by experimental data, artificial neural network is trained.
simulated annealing algorithm are presented in Trained network predicts the flash width, flash height
Table 8. It presents the optimized values of flash and flash thickness more closely. The percentage
features for minimized metal loss. The corresponding variation between the actual and predicted is around
input parameters are also present in the same table. 1.5%. The global optimization technique called
Table 4—Experimental, predicted and percentage of errors for FW, FH and FT with HP
Sl.No. HP Heating pressure versus flash parameters
Experimental value (mm) Predicted value (mm) Percentage of errors
EFW EFH EFT PFW PFH PFT ∆FW ∆FH ∆FT
1 22 13.9 9.35 6.47 14.0804 9.3963 6.5643 -1.2978 -0.4952 -1.4575
2 23 13.65 9.4 6.82 13.6972 9.4571 6.7849 -0.3458 -0.6074 0.5147
3 24 13.95 9.45 6.9 13.9362 9.5637 6.881 0.0989 -1.2032 0.2754
4 25 14.1 9.8 6.95 14.0527 9.6512 6.9827 0.3355 1.5184 -0.4705

Table 5—Experimental, predicted and percentage of errors for FW, FH and FT with HT

Sl. No. HT Heating time versus flash parameters


Experimental value (mm) Predicted value (mm) Percentage of errors
EFW EFH EFT PFW PFH PFT ∆FW ∆FH ∆FT
1 8 13.9 9.35 6.47 14.0804 9.3963 6.5643 -1.2978 -0.4952 -1.4575
2 9 13.65 9.4 6.82 13.6972 9.4571 6.7849 -0.3458 -0.6074 0.5147
3 9 13.95 9.45 6.9 13.9362 9.5637 6.881 0.0989 -1.2032 0.2754
4 10 14.1 9.8 6.95 14.0527 9.6512 6.9827 0.3355 1.5184 -0.4705

Table 6—Experimental, predicted and percentage of errors for FW, FH and FT with UP

Sl.No. UP Upsetting pressure verses flash parameters


Experimental value (mm) Predicted value (mm) Percentage of errors
EFW EFH EFT PFW PFH PFT ∆FW ∆FH ∆FT
1 42 13.9 9.35 6.47 14.0804 9.3963 6.5643 -1.2978 -0.4952 -1.4575
2 43 13.65 9.4 6.82 13.6972 9.4571 6.7849 -0.3458 -0.6074 0.5147
3 44 13.95 9.45 6.9 13.9362 9.5637 6.881 0.0989 -1.2032 0.2754
4 45 14.1 9.8 6.95 14.0527 9.6512 6.9827 0.3355 1.5184 -0.4705

Table 7—Experimental, predicted and percentage of errors for FW, FH and FT with UT

Sl.No. UT Upsetting time versus flash parameters


Experimental value (mm) Predicted value (mm) Percentage of errors
EFW EFH EFT PFW PFH PFT ∆FW ∆FH ∆FT
1 6 13.9 9.35 6.47 14.0804 9.3963 6.5643 -1.2978 -0.4952 -1.4575
2 7 13.65 9.4 6.82 13.6972 9.4571 6.7849 -0.3458 -0.6074 0.5147
3 7 13.95 9.45 6.9 13.9362 9.5637 6.881 0.0989 -1.2032 0.2754
4 7 14.1 9.8 6.95 14.0527 9.6512 6.9827 0.3355 1.5184 -0.4705
SATHIYA et al.: OPTIMIZATION OF FRICTION WELDING PARAMETERS 43

Fig. 7—Experimental and theoretical variation of flash features


with heating pressure Fig. 9—Experimental and theoretical variation of flash features
with upsetting pressure

Fig. 8—Experimental and theoretical variation of flash features Fig. 10—Experimental and theoretical variation of flash features
with heating time with upsetting time
44 INDIAN J. ENG. MATER. SCI., FEBRUARY 2006

Table 8—Optimized parameters for minimum metal loss for friction weldments

Heating pressure Heating time Upsetting pressure Upsetting time Flash width Flash height Flash thickness
(bar) (s) (bar) (s) (mm) (mm) (mm)
20 3 40 5 6.5076 7.3018 3.7026

simulated annealing is applied to the network model UTU = upper bounds of upsetting time
to achieve optimized parameters. Through the TSL = lower bounds of tensile strength
TSU = upper bounds of tensile strength
optimized input parameters, minimized flash features x1 = initial point
such as flash width (6.5 mm), flash height (7.3 mm) x2 = second point
and flash thickness (3.70 mm) can be obtained. Ts = initial temperature
Te = final temperature
Cr = cooling rate
Nomenclature
σ = variance
HP = heating pressure
n = number of random numbers
HT = heating time
ri = random numbers
UP = upsetting pressure
f = objective function
UT = upsetting time
FW = flash width
FH = flash height References
FT = flash thickness 1 Chang U I & Frish J, Weld J, (1974) 24s-35s.
TS = tensile strength 2 Graville B A & Read J A, Weld J, (1974) 161s-169s.
EFW = experimental flash width 3 Zhang Y M, Kovacevic R & Li L, Int J Machine Tools &
EFH = experimental flash height Manuf, l36 (7) (1996) 799-816.
EFT = experimental flash thickness 4 Elena Koleva, VACUUM-Surf Engi, Surf Instrument Vacuum
PFW = predicted flash width Technol, 162 (2001) 151-157.
PFH = predicted flash height 5 Tarng Y S, Tsai H L & Yeh S S, Int J Machine Tools &
PFT = predicted flash thickness Manuf, 39 (1999) 1427-1438.
∆FW = percentage of error between experimental flash width 6 Metropolis N, Rosenbluth A, Rosanblush M, Teller A &
and predicted flash width Teller E, J Chem Phys, 21 (1953) 1087-1092.
∆FH = percentage of error between experimental flash height 7 Kirkpatrik S, Gelatt C D & Vacchi M P, Science, 220 (4958)
and predicted flash height (1983) 671-680.
∆FT = percentage of error between experimental flash 8 Laarhoven P J M & Aarts E H L, Simulated Annealing
thickness and predicted flash thickness Theory and Applications, (Kluwer Academic Publishers,
ANN = artificial neural network Landon), 1989.
W1 = weights for the normalized flash width [FW] 9 Ruterbar R A, EEE Circuits Devices Mag, 5 (1) (1989)
W2 = weights for the normalized flash height [FH] 19-26.
W3 = weights for the normalized flash thickness [FT] 10 Lee B W & Shan B J, Hardware Annealing in Analog VLS,
HPL = lower bounds of heating pressure Neuro computing, (Kluwar Academic Publishers, Landon),
HPU = upper bounds of heating pressure 1991.
UPL = lower bounds of upsetting pressure 11 Kalyanmoy Deb, Optimizations for engineering design-
UPU = upper bounds of upsetting pressure Algorithm and examples (Prentice-Hall of India, Private Ltd.
UTL = lower bounds of upsetting time New Delhi), 1996, 320.

You might also like