Beruflich Dokumente
Kultur Dokumente
welding processes
T. W. Liao* and S. Daftardar
This paper proposes a model based approach for the optimisation of friction stir welding
processes. The proposed approach starts with building a model of the process. For this study, the
thermal model developed by Chao and his associate for friction stir welding of AA 2195-T8 is
replicated using Fluent. Once developed, the thermal model is then used to simulate the process.
Two surrogate models, one linear and one non-linear, are constructed to relate three process
parameters with maximum temperature at a selected location, using the simulation data
generated by the thermal model. A constrained optimisation model is next formulated, which is
eventually solved by five population based metaheuristrics to find the optimal solutions for the
studied friction stir welding process. The optimal solutions are primarily constrained by the lower
bound of the temperature. The lower this lower bound temperature is the higher travel speed can
go. The linear surrogate model results in a slightly better optimal solution than the non-linear
model when the temperature constraint is loose and the converse is true when the temperature
constraint is tight. Comparing all five metaheuristics, differential evolution has the best
performance, followed by particle swarm optimisation, ant colony optimisation, genetic algorithm,
and lastly harmony search.
Keywords: Friction stir welding, Thermal model, Metaheuristic, Ant colony optimisation, Differential evolution, Genetic algorithm, Harmony search, Particle
swarm optimisation, Finite volume
(ii) bounds of process parameter values. search and particle swarm optimisation. The specific
In welding, throughput for some duration of time can be algorithms implemented or toolbox used are the ACOR
measured by the length of weld completed that depends algorithm proposed by Socha and Dorigo (2008), the
much on the welding speed used. Without energy input, differential evolution (DE) algorithm proposed by Storn
welds cannot be made. Given that equipment cost and and Price (1997), the GA toolbox developed by Houck
labour cost are fixed, energy input is the dominant cost et al. (1995), the improved harmony search (IHS)
component. Since weld quality is the result of thermal algorithm developed by Mahdavi et al. (2007) and the
history during welding, the weld quality constraint can adaptive PSO algorithm proposed by Iwasaki et al.
be equated with a temperature constraint. Of course, the (2006). To handle constraints, the parameterless penalty
possible values of a process parameter are often limited method proposed by Deb (2000) is incorporated into
to the range available in the system used to carry out the each metaheuristic mentioned above. The resultant
process. This range can be further reduced as knowledge algorithms are called ACORz, DEz, GAz, IHSz
in operating the system for a particular job is gained and APSOz respectively.
over time. Major parameters associated with the ACORz
In this study, the two conflicting objectives are algorithm include number of ants n, maximum number
handled by combing them into one single objective of function evaluations maxnfe, a parameter that
function. The weights applied to each of the two controls intensification v. diversification q and a positive
objectives are assumed equal. To offset the magnitude parameter that has an effect similar to that of the
difference between the two objectives, the objective with pheromone evaporation rate j, the size of solution
lower magnitude is multiplied by a constant of 500. archive T. For this study, they are fixed at 30, 150 000,
Except GA, all metaheuristics attempt to minimise the 0?7, 0?7 and 15 respectively. Major parameters asso-
aggregated objective. Specifically, two optimisation ciated with the DEz algorithm include number of
models to be solved by the five population based parameter vectors n, maximum number of function
metaheuristics have the following form evaluations maxnfe, a parameter [0, 2] that controls the
amplification of differential variation F, and a crossover
Min(H{500S)½or Max({Hz500S) in the case of GA parameter [0, 1] that is used to construct a trial vector,
CR. For this study, they are fixed at 30, 150 000, 0?5 and
0?95 respectively. Major parameters associated with the
s.t. TLB(T(TUB
GAz algorithm include number of parameter vectors in
1500(H(1860
the population n and maximum number of function
1?5(S(4?2
evaluations maxnfe. All other GA parameters are taken
19(D(25?4
as the default values. Major parameters associated with
The two models differ primarily on the equation for T: the IHSz algorithm include harmony memory size n,
called model 1 if equation (3) is used and model 2 if maximum number of function evaluations (improvisa-
equation (4) is used instead, for easy reference later. The tions) maxnfe, harmony memory considering rate
bounds for T will be varied to investigate the effect HMCR, maximum and minimum pitch adjusting rates
of setting a tighter bound on the optimal solution and PARmax and PARmin, and maximum and minimum
the optimal objective value in the section on ‘Results’. bandwidths, bwmax and bwmin. In this study, they are
The bounds for the three process variables are taken as fixed at 30, 150 000, 0?9, 0?99, 0?45, 4 and 0?00001
the lowest and highest experimented values in this study, respectively. Major parameters associated with the
as given in the last three constraints. APSOz algorithm include number of particles n,
maximum number of function evaluations maxnfe, two
Population based metaheuristics positive constants associated with the velocity update
equation c1 and c2, and six parameters associated with
Metaheuristics are high level strategies for exploring the adapting scheme wmax, wmin, wdelta, vstart, vinit and
search spaces using different methods. Many metaheur- Tend. For this study, they are fixed at 30, 150 000, 1?3,
istics have been proposed so far and there are different 1?3, 0?9, 0?4, 0?1, 2, 4 and 4000 respectively.
ways to classify them. Based on the number of solutions
used at the same time as the classification criterion,
metaheuristics can be grouped into two major cate- Results
gories: single point search metaheuristics and population Since the global optimal objective function value of the
based metaheuristics. Single point search metaheuristics targeted FSW process is unknown, a relatively high
include simulated annealing, tabu search, iterated local maximum number of total function evaluations of
search, variable neighbourhood search, etc. Population 150 000 was consistently used as the stopping criterion
based metaheuristics include ant colony optimisation, for each metaheuristic to prevent premature stopping of
differential evolution, genetic algorithm, harmony the search. Generally, the higher the population size, the
search, particle swarm optimisation, etc. Generally, better the chance to find the optimal solution. However,
population based metaheuristics are more effective for the computational cost is also higher with larger
constrained function optimisation problems than single population size. For fair comparison, each metaheuristic
point search metaheuristics. Population based meta- employs the same population size of 30. The parameter
heuristics are selected for this study because optimisa- values associated with each metaheuristic were set at the
tion of FSW processes is a constrained function values recommended by the original authors, as given in
optimisation problem. the previous section. Since all five metaheuristics are
Five population based metaheuristics were implemen- stochastic in nature, 30 runs were made to produce
ted for this study. They are ant colony optimisation, sufficient statistical data. The best, worst and median
differential evolution, genetic algorithm, harmony results of objective values and amounts of CPU time
taken for each metaheuristic were computed. The (iii) comparing ACORz with GAz, ACORz
solutions for the best result obtained by each metaheur- seems to have better overall performance in
istic were also recorded. All programs were coded in terms of objective value, CPU time taken and
Matlab and all experiments were carried out using Dell number of runs found the best solution (23 v.
Latitude D830 with Intel Core2 Dual CPU T9300 at 21) than GAz
2?5 GHz. (iv) the IHSz algorithm has the worst perfor-
mance. It seems not able to converge to the
Results of model 1 best solution. Nevertheless, the solution found
Model 1, the optimisation model that uses the fitted is very close to the best. In fact, the deviation
linear regression equation of T, i.e. equation (3), was away from the best solution might be insignif-
solved using a wider range of T, i.e. TLB5425 and icant because it might be less than the accuracy
TUB5475. This temperature range and bound values of the welding system to implement the solu-
should be set in consideration of material properties, tion. Two notable properties about the IHSz
particularly the melting temperature and its sensitivity to algorithm are fast and small variance in its
heat input variation. A material with higher sensitivity solutions.
to heat input would have a narrower workable To investigate the effect of narrowing the range of
temperature range. For example, AA 5083-O can have maximum temperature, model 1 is again solved using a
a wider temperature range than AA 2024-T6 and narrower range of T, i.e. TLB5450 and TUB5475. Table 3
AA 7075-T6 because it is not as sensitive to heat input summarises the results obtained by all five population
as reported by Nakata et al. (2000). The melting based metaheuristics. The results indicate that:
temperature of AA 2195-T8 is y540 uC. The authors (i) the best objective value is 558?4879, which is
are not aware of any study that investigated the range of achieved by four algorithms: ACORz, DEz,
temperature in AA 2195-T8 that led to good welds. GAz and APSOz. The optimal solution is
Further experiments are such needed to investigate the H51860 W, S52 ?603024 mm s21 and
sensitivity of AA 2195-T8 to heat input. To enable this D519 mm
study, the bound values selected are a rough guess (ii) both DEz and APSOz algorithms found the
around known good temperature value. best objective value in all 30 runs. However, the
Table 2 summarises the results obtained by all five DEz algorithm is more time consuming than
population based metaheuristics. The results indicate the APSOz algorithm
that: (iii) comparing ACORz with GAz, ACORz
(i) the best objective value is 306?4718, which is seems to have better overall performance in
achieved by four algorithms: ACORz, DEz, terms of objective value, CPU time taken and
GAz and APSOz. The optimal solution is number of runs found the best solution (27 v.
H51860 W, S53 ?107056 mm s21 and 22) than GAz
D519 mm (iv) the IHSz algorithm again has the worst
(ii) both DEz and APSOz algorithms found the performance. It seems not able to converge to
best objective value in all 30 runs. However, the the best solution. But the solution found might
DEz algorithm requires more CPU time than be good enough in practice because the devia-
the APSOz algorithm tion might be less than the accuracy of the
system implementing the solution. The proper- It is also fast and the variance of solutions
ties of fast and small variance are again found is relatively small
observed (v) with further increasing TLB, the welding speed is
(v) with increasing TLB, the lower bound of forced to further reduce in order to provide even
maximum temperature, the welding speed is more heat input, which in turn leads to an even
forced to reduce in order to provide more heat higher objective value.
input, which in turn leads to a higher objective
value.
Results of model 2
The above reduction in temperature range does not seem Just like model 1, model 2, the optimisation model that
to affect the search ability of any metaheuristic too uses the fitted non-linear regression equation of T, i.e.
much. From the physical viewpoint, the reduction in equation (4), was solved using a wider range of T, i.e.
temperature range means a tighter control of the TLB5425 and TUB5475. Table 5 summarises the results
thermal field, which shall lead to more consistent weld obtained by all five population based metaheuristics.
quality. From the search viewpoint, the reduction in The results indicate that:
temperature range means fewer feasible solutions, thus (i) the best objective value is 347?8579 with optimal
solution of H51860 W, S53?024284 mm s21
making it more difficult to find the optimal solution. To
and D519 mm which is achieved by four
show this effect, model 1 is again solved using an even
algorithms: ACO R z, DEz, GAz and
narrower range of T, i.e. TLB5470 and TUB5475.
APSOz. Note that this objective value is higher
Table 4 summarises the results obtained by all five
than that obtained using model 1 and the
population based metaheuristics. The results indicate
welding speed is lower. This result seems to
that:
indicate that optimisation on model 1 leads to a
(i) the best objective value is 760?1008, which is
better solution
achieved by four algorithms: ACORz, DEz, (ii) both DEz and APSOz algorithms found the
GAz and APSOz. The optimal solution is best objective value in all 30 runs. However, the
H51860 W, S52 ?199798 mm s21 and DEz algorithm requires more CPU time than
D519 mm the APSOz algorithm
(ii) DEz is the only algorithm that found the best (iii) comparing ACORz with GAz, ACORz
objective value in all 30 runs. APSOz found seems to have better overall performance in
the best objective value in 28 of 30 runs. The terms of objective value, CPU time taken and
search ability of the APSOz algorithm is number of runs found the best solution (26 v.
apparently reduced as the result of further 21) than GAz
reducing the temperature range (iv) the IHSz algorithm has the worst perfor-
(iii) comparing ACORz with GAz, ACORz mance. It seems not able to converge to the
seems to have better overall performance in best solution. Nevertheless, the solution found
terms of objective value, CPU time taken and is very close to the best. The deviation away
number of runs found the best solution (24 v. from the best solution might be insignificant if it
20) than GAz is less than the accuracy of the welding system
(iv) the IHSz algorithm again has the worst to implement the solution. It should be noted
performance. It seems not able to converge to that the IHSz algorithm is fast and its
the best solution, but close enough practically. solutions do not vary much.
Table 5 Optimisation results of model 2 with TLB5425 and TUB5475
To investigate the effect of narrowing the range of (i) the best objective value is 740?1079 with optimal
maximum temperature, model 2 is again solved using a solution of H51860 W, S52?239784 mm s21
narrower range of T, i.e. TLB5450 and TUB5475, just and D519 mm, which is achieved by four
like model 1. Table 6 summarises the results obtained by algorithms: ACO R z, DEz, GAz and
all five population based metaheuristics. The results APSOz. Note that unlike the previous two
indicate that: cases, this objective value is lower than that
(i) the best objective value is 570?6282 with optimal obtained using model 1 and the welding speed is
solution of H51860 W, S52?578744 mm s21 higher. This result seems to indicate that
and D519 mm, which is achieved by four optimisation on model 2 leads to a better
algorithms: ACO R z, DEz, GAz and solution in this particular case with the tightest
APSOz. Note that this objective value is higher temperature range
than that obtained using model 1 and the (ii) DEz is the only algorithm that found the best
welding speed is lower. This result seems to objective value in all 30 runs. APSOz found
indicate again that optimisation on model 1 the best objective value in 27 of 30 runs. The
leads to a better solution search ability of the APSOz is slightly com-
(ii) both DEz and APSOz algorithms found the promised as the result of further reducing the
best objective value in all 30 runs. However, the temperature range
DEz algorithm consumes more CPU time than (iii) comparing ACORz with GAz, ACORz
the APSOz algorithm seems to have better overall performance in
(iii) comparing ACORz with GAz, ACORz terms of objective value, CPU time taken and
seems to have better overall performance in number of runs found the best solution (23 v.
terms of objective value, CPU time taken and 17) than GAz
number of runs found the best solution (27 v. (iv) the IHSz algorithm again has the worst
22) than GAz performance. It seems not able to converge to
(iv) the IHSz algorithm again has the worst the best solution. The deviation of 0?111 W in
performance. It seems not able to converge to heat input and ,0?001 in welding speed away
the best solution. But it can be argued again from the best solution might be insignificant in
that this deviation might be insignificant in practice. The HISz algorithm is the fastest
practice among all five metaheuristics. The variance of
(v) with increasing TLB, the lower bound of max- solutions found is relatively small
imum temperature, the welding speed is forced to (v) with further increasing TLB, the welding speed is
reduce in order to provide more heat input, forced to further reduce in order to provide even
which in turn leads to a higher objective value. more heat input, which in turn leads to an even
Just like model 1, the above reduction in temperature higher objective value.
range does not seem to affect the search ability of each
metaheuristic too much. Therefore, model 2 is again
solved using an even narrower range of T, i.e. TLB5470
Discussion
and TUB5475. Table 7 summarises the results obtained There have been a few studies that focused on the
by all five population based metaheuristics. The results friction stir welding of AA 2195-T8. Seidel and
indicate that: Reynolds (2001) investigated the material flow using a
marker insertion technique. Kroninger and Reynolds found is close enough. It can be argued that the
(2002) examined the fracture resistance of friction stir deviation is insignificant because it might be less than
welds in AA 2195-T8. Prangnell and Heason (2005) used the accuracy of the physical system used to implement
the ‘stop action technique’ to observe grain structure the solution.
formation during FSW of AA 2195-T8. Schneider et al. In terms of CPU time, the IHSz requires the least
(2005) attempted to correlate the microstructure with and the remaining four metaheuristics can be roughly
the variation in thermal history the material experiences ranked in ascending order as follows: ACORz, GAz,
during the FSW process. Fonda and Bingert (2006) APSOz and DEz. Note that this order is nearly
analysed the microstructure to reveal precipitation opposite to the ranking order obtained based on
processes, grain evolution mechanisms and crystallo- objective values. Since process optimisation is often
graphic texture within the weld. This study is the first carried out offline, CPU time requirement is not
attempt to optimise the process of FSW of AA 2195-T8. necessarily a major concern. Therefore, ranking based
Note that the corresponding temperature values of all on objective values is favoured in this study.
of the best solutions obtained in all cases tested in the To further investigate how many function evaluations
section on ‘Results’ take on the lower bounds of the taken by DEz and APSOz respectively to find the best
temperature constraint. This implies that weighting more solution, an additional criterion is used as the stopping
heavily on the objective of maximising the throughput criterion other than the maximum number of function
than the objective of minimising heat input, instead of evaluation. This criterion compares the best solution
equally weighted, will not change the best solutions at all. found so far during search with the global optimum,
This observation was confirmed with experimental runs. which is set to be the best solutions reported in the
Hence, to maximise productivity, the lower bound of section on ‘Results’. The stopping criterion is triggered
temperature thus should be set as low as possible to the when the best solution found so far is less than the
point that good weld can be made. The three lower bound global optimum. Both models 1 and 2 with the wider
values used in this study are 425, 450 and 470 uC, which temperature range, TLB5425 and TUB5475, are solved
are about 79, 83 and 88% of the melting temperature by DEz and APSOz one more time for this investiga-
(,540uC). It seems that a ,425uC lower bound could be tion. Table 8 summarises the results. The results indicate
set to allow the use of higher travel speed. But further that:
experiments will be needed to confirm this. (i) on average, DEz takes y4,000 function evalua-
Comparing the two surrogate models, the results tions and 1?8 s CPU time whereas APSOz takes
indicate that optimisation on model 1 produces better y1500 function evaluations and 0?4 s to con-
results when the temperature range is loose whereas verge to the assumed global optimum. DEz thus
optimisation on model 2 produces better results when requires higher CPU time and converges slower
the temperature range is tight. This statement holds true than APSOz. The maximum number of function
regardless which metaheuristic is used. evaluation of 150 000 used in the previous section
The test results presented in the previous section is apparently an overkill for both algorithms
clearly indicate that in terms of objective value, the (ii) owing to its stochastic nature, the DEz algo-
DEz algorithm has the best performance and the rithm did not find the global optimum in all 30
remaining four metaheuristics can be ranked in runs this time, but only 29 times.
the following order: APSOz, ACORz, GAz and It should be pointed out that the proposed model based
IHSz. Even though losing its search ability a bit when optimisation approach is general in nature. Despite of
the tightest temperature range is used, the APSOz the authors’ choice of Chao’s model, other thermal
algorithm found as many good solutions as DEz and in models developed for FSW processes by Gould and
less CPU time. Therefore, APSOz seems to be a strong Feng (1998), McClure et al. (1998), Chao and Qi (1998),
contender as well. Though not as good as DEz and Khandkar et al. (2003), Simar et al. (2006), Vilaça et al.
APSOz, both the ACORz and GAz algorithms (2005, 2007), Colegrove et al. (2007), Hamilton et al.
found the best solutions in all cases, for more than half (2008) and Schmidt and Hattel (2008) can also be
of the runs. Therefore, both the ACORz and GAz optimised similarly. Actually, more sophisticated models
algorithms are also good metaheuristics to use with the such as coupled or uncoupled thermomechanic models
ACORz algorithm slightly better than the GAz or models considering material flow and microstructural
algorithm. As far as IHSz algorithm is concerned, it formation can also be optimised. The main differences
does not converge to the best solution found by the lie in the number of variables and constraints involved
other four metaheuristics. However, the best solution it and may be in the objective functions, too.
Table 8 Optimisation results with assumed global optimum as additional stopping criterion
DEz APSOz
33. P. Vilaça, L. Quintino and J. F. dos Santos: J. Mater. Process. 34. P. Vilaça, L. Quintino, J. F. dos Santos, R. Zettler and S. Sheikhi:
Technol., 2005, 169, 452–465. Mater. Sci. Eng. A, 2007, A445–A446, 501–508.