Professional Documents
Culture Documents
Optimizing Drilling Induced Delamination in GFRP Composites Using Genetic Algorithm & Particle Swarm Optimisation
Optimizing Drilling Induced Delamination in GFRP Composites Using Genetic Algorithm & Particle Swarm Optimisation
ABSTRACT
Composites are widely used in several applications ranging from automotive to aircraft industry due to their
high strength to weight ratio. More often than not drilling on these composite laminates are conducted to serve
some functional or aesthetic requirement. Delamination caused due to drilling pose a severe problem to the
integrity of the structure. It is often not possible to develop an exact mathematical model to predict the delami-
nation associated with such drilling. So, in this paper, an empirical model is developed based on the extensive
experiments performed on polyester composite reinforced with chopped fibreglass. To account for the various
parameters a Box-Behnken design of experiments is conducted for four parameters (material thickness, drill
diameter, spindle speed, and feed rate) each having threedistinct levels. Genetic Algorithm (GA) and Particle
Swarm Optimization (PSO) techniques are then used for predicting the global optimum (minimum delamina-
tion factor). The performance of both GA and PSO in terms of predicting the global optimum is found to be
same. However, PSO converged much faster and required far lesser computational time.
Keywords: Box-Behnken design, Drilling, GFRP composite, Response surface methodology (RSM), GA
Fig.
Fig.2: Normal probability
2. Normal plot of plot
probability residuals for delami-for Fig.
of residuals Fig.3: Residuals vs predicted
3. Residuals response forresponse
vs predicted delamination
for
nation factor factor
delamination factor delamination factor
Advanced Composites Letters, Vol. 27, Iss. 1, 2018 5
181 contradiction is perhaps the fact that the +1 level, i.e. the higher level considered by Srinivasan et al.[22] is 5000
K.Kalita,
182P. K.
rpmMallick, A.K.
whereas in the Bhoi,
current study R.
the K.Ghadai
upper level is at merely 1100 rpm. Also, it is worth mentioning here that
183 Srinivasan et al.[22] had used a polypropylene matrix as compared to thepolyester matrix utilized in this study.
Fig. 4. Interaction effect of spindle speed (N) and material thickness (t) on delamination factor
Fig. at5.different drilleffect
Interaction diameters
of spindle speed (N) and feed rate (f) on delamination factor at different drill diameters and material
and feed rates.
Fig. 4: Interaction effect of spindle speed (N) and mate-
thickness.
Fig. 5: Interaction effect of spindle speed (N) and feed
161 rial thickness (t) on delamination factor at different
Fig. 4show the interaction effect of material thickness and spindle speed on delamination
184 factor at3.2drill
various levels of of (f)
Performance
rate on delamination factor at different drill diameters
the mathematical model
162 diameters
feed rate and drill diameter. In general, delamination and
factor feedwith
increases rates.
increase185
in feed rate. Behera et al. [18] and material
Before proceeding to optimise the process parameters involved thickness.
in the drilling of GFRP composites by a genetic
163 and Rubio et al. [19] has also reported a similar trend. This isbecause higher feed rate
186induces higher thrust
algorithm forces,
or particle swarm optimisation, it is essential to verify the performance of the selected RSM model. Fig. 6
164 thereby causing more delamination. Also, the material removal rate and the contact187 surface shows
increases
theatcomparison
the higher of the predicted delamination factor calculated using Eqn. 15 and the experimental output. It
165
Fig. 5 show that the effect of feed rate and spindle 4.2 Performance of the mathematical model
feed which is responsible for the additional delamination. It is also seen that the delamination
188 is increases
seen that as
thethe drill model predicts the delamination factor in the drilling of GFRP composites with very high
present
166 speed at different levels of material thickness and Before proceeding to optimise the process param-
diameter are increased, given that all other parameters are kept same. The greater contact
189 and accuracy. The mostdue
larger thrust force variation (2.5%) in the predicted and the measured values is seen in the trial no. 9. Other than
167 drillarediameter.
to larger drill diameter In general, delamination
responsible for this. this particular caseeters
190 decreases the modelinvolved
is quiteaccurate,in
andthe drilling
an overall of GFRP
average variation composites
of only 0.86% is seen.
168
with increase in material thickness. At lower spindle by a genetic algorithm or particle swarm optimisa-
Fig. 5 show that the effect of feed rate and spindle speed at different levels of material thickness and drill diameter.
169 In general, delamination decreases with increase in material thickness. At lower spindle speeds, the change in drill
170
speeds, the change in drill diameter has anegligible tion, it is essential to verify the performance of the
diameter has anegligible effect on the delamination factor. However, as the spindle speed is increased, the increase
171 in drill diametereffect ondelamination
enhances the the delamination factor.
in the composite. Overall from allHowever, is evident selected
asit the
the interaction plots, that RSM model. Fig. 6 shows the comparison
172 at alow feed rate spindle speed
and low spindle speed,is increased,
delamination is lesser.the increase
This is in drill
in direct agreement di-
with the of the predicted delamination factor calculated using
conclusions
173 ameter enhances the delamination in the composite. Eqn.
ofBehera et al. [18]; Vankanti and Ganta [20]; and Hansda and Banerjee [21]. Vankanti and Ganta [20]have used an 15 and the experimental output. It is seen that
174
Overall from all the interaction plots, it is evident the present model predicts the delamination factor
L9 orthogonal array for thedesign of the experiments and used an ANOVA test to determine the significance of the
175 parameters. Hansda and Banerjee [21] had also used a similar L9 orthogonal array to arrive at the same conclusion.
176 thatalsoatreported
Latha et al. [7]have alow feed rate
that combination andspeed
of spindle lowand spindle
feed rate also speed, de- inin the drilling of GFRP composites with very high
affects the delamination
177 drilling of GFRP lamination is lesser. This is in direct agreement with
composites.The delamination can be further reduced by combining the low feed rate and accuracy.
low The most variation (2.5%) in the predicted
178 the conclusions of Behera et al. [18]; Vankanti and and the measured values is seen in the trial no. 9.
spindle speed settings with a higher laminate thickness. Srinivasan et al.[22]have reported that a higher spindle
179 speed is beneficial in reducing the delamination as it softens the polymer due to high tool traverse. However, in the
Ganta [20]; and Hansda and Banerjee [21]. Vankanti 9|Page Other than this particular case the model is quite-
180 current study delamination is found to be less at a lower spindle speed. One of the possible reason for this
and Ganta [20] have used an L9 orthogonal array for accurate, and an overall average variation of only
8 | P a g e thedesign of the experiments and used an ANOVA
0.86% is seen.
test to determine the significance of the parameters.
Hansda and Banerjee [21] had also used a similar
L9 orthogonal array to arrive at the same conclu-
sion. Latha et al. [7] have also reported that com-
bination of spindle speed and feed rate also affects
the delamination in drilling of GFRP composites.
The delamination can be further reduced by com-
bining the low feed rate and low spindle speed set-
tings with a higher laminate thickness. Srinivasan et
al.[22] have reported that a higher spindle speed is
beneficial in reducing the delamination as it softens
the polymer due to high tool traverse. However, in
the current study delamination is found to be less at
a lower spindle speed. One of the possible reason for
this contradiction is perhaps the fact that the +1 lev-
el, i.e. the higher level considered by Srinivasan et Fig. 6. Comparison of the experimental and predicted results
al.[22] is 5000 rpm whereas in the current study the Fig. 6: Comparison of the experimental and predicted
upper level is at merely 1100 rpm. Also, it is worth results
mentioning here that Srinivasan et al.[22] had used
a polypropylene matrix as compared to thepolyester 4.3 Predicting optimal process parameters by GA
matrix utilized in this study. The mathematical model for delamination factor in
the drilling of GFRP composites formed using the
Fig. 7. Performance of the GA across 100 generations Fig.8. Performance of the PSO vs GA across 100 generations
Fig. 7: Performance of the GA across 100 generations Fig.8: Performance of the PSO vs GA across 100 gen-
191 3.3 Predicting the optimal process parameters by GA
erations
Box-Behnken RSM method is192 used asThethe objective
mathematical modelthe two. In here
for delamination factorthe best
in the solution
drilling of GFRP from 1st genera-
composites formed using the Box-Behnk
function for optimisation using a genetic
193 algorithm.
RSM method is used astion is replaced
the objective functionby for a better solution
optimisation using found
a geneticinalgorithm.
the 2nd A FORTRAN code
A FORTRAN code is compiled 194 forcompiled
executing the generation,
for executing the GA. The problemand the process as
is formulated goes on until a optimisation
an unconstrained global problem where t
GA. The problem is formulated 195as anobjective
unconstrained best is achieved
is to minimise delamination factor. in the 23 rd
generation. The optimum
optimisation problem where the 196 objective
i.e. minimiseto
is mi- combinations
delamination factor, with the of parameters predicted by the GA are
limits,
nimise delamination factor. reported in Table 4 along with the predicted mini-
i.e. minimise delamination factor, with the limits, mum delamination factor.
8 ≤ t ≤16
10 ≤ D ≤14 4.4 Predicting optimal process parameters by PSO
400 ≤ N ≤1100 Another FORTRAN code iscompiled to implement
Page 197 The various GA parameters considered in the runSwarm are-chromosome length 16, Population
0.1 ≤ f ≤0.275 the PSO technique. size plays a crucial role 100, Generations 10
The various GA parameters considered in the run in PSO, and robustness and computation cost of the
198 Crossover rate 0.90 and Mutation rate 0.02. For the current problem, several combinations of crossover an
are-chromo some length 16, 199 Population
mutation100, Gen- were
probabilities algorithm
tried andisthis also affected
particular by it. was
combination Small foundpopulation
to be the best in augmenting th
erations 100, Crossover rate 0.90 and Mutation rate
200 explorative and size
exploitative may
traits of result
the GA. in
The local
high convergence;
crossover allows thelarge
GA to size in- the ‘good structures’
maintain
0.02. For the current problem,201 several‘fitcombinations
individuals’ while the creases computational
low mutation ensures that keyefforts and are
genetic traits may make
not lost whileslow
flipping a bit randomly. Fi
of crossover and mutation probabilities
202 were tried convergence [23]. Thus, in the present study, a par-
7illustrates the performance of the GA across the generations. Though at the end of generation 1 the average fitne
and this particular combination 203was isfound
seen totobebe ticlethan
the higher
a little sizetheof
best100,
fitnessinertia weight
in the same linearly
generation, the decreasing
average fitness sharply drops till 2
best in augmenting the explorative
204 and exploitative from 0.90 to 0.40
generation, after which it zig-zags around the best fitness
and c =c
1 line.
2
=2 is considered. The
This is because since the GA randomly creates
traits of the GA. The high crossover
205 allows the GA PSO is allowed to run for 100 generations to make
population of 100 individuals, it is quitepossible to have some individuals much superior to the rest of t
a
to maintain the ‘good structures’ in ‘fit individu- meaningful comparison with the GA in terms of ac-
206 population. As the generations progress, the GA compares ‘the best so far’ individual to the best of the curre
als’ while the low mutation ensures that key genetic curacy and computational speed. It is seen from Fig.
207 generation and continue to carry forward the better of the two. In here the best solution from 1st generation
traits are not lost while flipping a bit randomly. Fig. 8 that the PSOnd algorithm rapidly converges to the
208 replaced by a better solution found in the 2 generation, and the process goes on until a global best is achieved in t
7 illustrates the performance of the GA across the global optimum in this case. Though the same glob-
209 23rd generation. The optimum combinations i.e.oftheparameters predicted by the GA arefactor
reportedisin Table4 along with t
generations. Though at the end of generation 1 the al optimum, minimum delamination
210
average fitness is seen to be a little higher than the predicted by both the GA and the PSO algorithm, it
predicted maximum delamination factor.
best fitness in the same generation, the average fit- is seen that the PSO issomewhat faster in locating
ness sharply drops till 25 generation,11after | P a which
ge it the optimum. The GA located the global optimum
zig-zags around the best fitness line. This is because in the 23rd generation whereas the PSO was able to
since the GA randomly creates a population of 100 attain this in its 16th solution. Moreover, the compu-
individuals, it is quite possible to have some indi- tational time for optimised response in case of PSO
viduals much superior to the rest of the population. (0.72 s) was found to be lesser than the GA (3.25
As the generations progress, the GA compares ‘the s). Confirmation experiments are also carried out to
best so far’ individual to the best of the current gen- validate the GA & PSO predictions and are reported
eration and continue to carry forward the better of in Table 4.
Table 4: Optimum parameters predicted by GA and PSO