Documente Academic
Documente Profesional
Documente Cultură
DOI 10.1007/s00170-016-9183-2
ORIGINAL ARTICLE
Received: 25 April 2016 / Accepted: 13 July 2016 / Published online: 26 July 2016
# Springer-Verlag London 2016
Abstract Incremental sheet forming (ISF) has emerged as each of the four inputs. It was found that sheet thickness and
one of the rapid prototyping processes in forming 3D-shaped wall angle influence the most the deformation energy and
products from design data without much need for human and unclamped geometrical accuracy, respectively. The findings
tool intervention. Past studies show that an extensive focus from this analysis are useful in monitoring deformation energy
has been paid on the analysis of surface roughness, clamped and thus promoting greener environment.
accuracy, and energy consumption in terms of machine stand-
by and positioning energy, while very little focus was paid to Keywords Incremental sheet forming . Optimization .
deformation energy and unclamped accuracy. In this context, System identification . Deformation energy . Genetic
system identification-based methods can play a vital role in programming
determining and evaluating hidden relationships between the
inputs (tool diameter, wall angle, sheet thickness, and step
down) and two characteristics, namely, deformation energy 1 Introduction
and unclamped accuracy. The present work proposes compre-
hensive study based on the feedback-evolutionary system The incremental sheet forming (ISF) process is an emerging
identification (FB-ESI) approach in predictive evaluation of rapid prototyping process for making complex and custom-
multiresponse characteristics. The validation and robustness ized parts by forming sheet metals based on progressive local-
of the models were done based on statistical error metrics, ized plastic deformation using a hemispherical headed tool on
cross-validation, new complexity metrics, and hypothesis computer numerically controlled machine tools (Fig. 1) [1].
tests. Furthermore, the model analysis based on the parametric The greater flexibility in its forming limits makes it a viable
and sensitivity procedures resulted in 2D and 3D plots that process for low-cost and small-batch production. ISF process
measure the nature of the two characteristics with respect to build the desired geometry of the designed part by accurate
tool path generation from computer-aided manufacturing soft-
ware without much need of human and tool intervention. The
* B. N. Panda advantages of this process include enhanced formability, low-
biranchi.panda3@gmail.com er forming forces, and greater flexibility in the process [2].
However, the wider applications of this technology and its
1
Department of Mechatronics Engineering, Shantou University,
commercialization is hindered by limitations in its geometrical
Shantou 515063, China accuracy, processing times, and higher consumption of energy
2
The State Key Laboratory of Digital Manufacturing Equipment and
needed for deforming metal sheets [3, 4]. These multiresponse
Technology, Huazhong University of Science and Technology, 1037 characteristics are highly influenced by technological param-
Luoyu Road, Wuhan 430074, China eters such as wall angle, step depth, tool diameter, spindle
3
IDMEC, Instituto Superior Tcnico, Universidade de Lisboa, Av. speed, and sheet thickness etc. [4].
Rovisco Pais, 1040-001 Lisboa, Portugal Significant attention on experimental work and numerical-
4
Veer Surendra Sai University of Technology, based finite element models have been paid in optimization of
Burla 768018, Sambalpur, India surface roughness, formability limits, and clamped geometric
1354 Int J Adv Manuf Technol (2017) 89:13531365
Wall angle
Work piece
Initial sheet
ss
thickness
ne
ck
hi
Initial sheet
t
et
he
ls
Step down
na
Fi
Experimental rig
accuracy of the formed part from ISF [5, 6]. Various variants In context of optimization of energy consumption of
of surface roughness such as mean surface roughness, root the ISF process, Duflou [13] suggested three general ways
mean square (RMS) surface roughness, and mean peak to such as the redesign of machine tools and control parts,
valley height were investigated based on the spindle speed allocation of the machine tool at reasonable level, and the
and step depth by formulation of an empirical model [7]. selection of the appropriate inputs (tool diameter, wall
Empirical models were formulated to study the effects of tool angle, sheet thickness, step size, etc.). Dittrich et al. [14]
radius and step depth on the surface roughness and formability presented the notion (exergy) of extracting the maximum
of pyramidal and conical shaped ISF-formed parts [8]. valuable work from the ISF process at a given state and
Formability was effectively measured by Ham and Jeswiet concluded that the material of the sheet used has
[9] using the design of experiment methodology based on
two inputs: maximum forming angle and effective strains.
The geometric accuracy of the ISF-formed parts is catego- Surface roughness,
rized into three components such as clamped accuracy, clamped geometrical
unclamped accuracy, and final accuracy (Fig. 2). A major accuracy and
emphasis was paid in improving the clamped accuracy of formability
the formed parts [3]. The works comprising of numerical Response
modeling was undertaken by Essa and Hartley [10] to mini- characteristics Direction
mize the sheet bending, springback, and pillow effects. The of ISF process towards
effect of the two different depth values and springback effect Motivation
on the shape deviations of the ISF-formed pyramid parts were
evaluated numerically. It was found that the strains arising due Deformation energy
to structural bending results in the deviations in final and Unclamped
geometrical accuracy
shape, while the springback has little effect. Statistical
modeling based on the design of the experiment methods
such as Box-Behnken were undertaken to study the ef- Statistical and
fects of the inputs (tool radius, step size, wall angle, sheet Experimental
thickness, and final product depth) on the geometric ac- Strategies
SI methods
curacy of the truncated cone formed from the ISF process
for ISF
[11]. Sheet thickness and step size were found to be the process
main inputs affecting accuracy. Similarly, Ambrogio et al.
[12] applied the design of an experiment methodology to
study the effect of these inputs on the dimensional accu- System
racy of the ISF-formed parts. Based on the qualitative Identification (SI)
analysis, it was found that the geometric accuracy was Theory
found to be highest at the bottom and the overall accuracy Fig. 2 Motivation behind choosing response characteristics and the
was 1 mm. system identification theory
Int J Adv Manuf Technol (2017) 89:13531365 1355
significant contribution to the total exergy input. A de- energy consumption of eight times can be achieved. Study
tailed quantitative analysis was performed by Branker on the effect of process parameters (sheet material, step
et al. [15] to measure the effects of feed rate, lubricant, down, and tool rotation speed at the two designed levels)
and step-down size on the carbon dioxide emission, cost, on energy consumption was evaluated [19]. The results
and energy consumption of the aluminum-designed hat concluded that the rotation speed of the tool influences
from the ISF process. His analysis concluded that at the energy consumption the most.
higher values of feed rate and step-down size and using Literature studies have been summarized in Fig. 2. It shows
the lubricant at a given state of the ISF process, the ener- that the extensive focus was given to optimization of surface
gy consumption was reduced by 69 % and cost by 8.4 %. roughness, clamped accuracy, energy consumption in terms of
Ingarao et al. [16] presented detailed empirical compari- machine standby, and positioning energy, while very little focus
sons in the context of energy consumption and material was paid to deformation energy and unclamped accuracy [2].
savings for the ISF and conventional stamping processes. Deformation energy is of prime importance because higher
The findings from the analysis reported that the ISF pro- deformation energy implies higher heat generation between
cess results in higher consumption of deformation energy the tool and sheet resulting in tool and part wear, and thus,
than stamping process does, while the material savings premature failure of the part. On the other hand, the attention
were more in the case of the ISF process. It was also was paid to improving the clamped accuracy with less focus
concluded that the energy consumed during forming by on unclamped accuracy, which is due to the development of
the ISF process exhibited higher dependence on the ma- stresses due to plastic deformation. The simultaneous optimi-
chine type, material type, part geometry shape, and thick- zation of both the deformation energy and the unclamped
ness. Further investigation by Ingarao et al. [17] on the geometric accuracy is a challenging multiobjective optimiza-
material type for energy consumption on the three differ- tion problem with conflicting objectives. Accurate geometric
ent machines (CNC milling machine, six-axis robot, and accuracy requires higher deformation energy and thus in-
the AMINO) was conducted. It was found that the mate- creases the cost and poses productivity and environmental
rial type has no influence on the energy consumption as- threat. The evolutionary system identification (E-SI) theory
pect of the ISF process on CNC milling machine, while [2024] specifically tasked in the automation of the formation
the six-axis robot was found to be influenced the most by of models from the given data can be applied. Genetic pro-
it. The AMINO setup was proved to be most efficient in gramming (GP) is one such approach that is based on evolu-
terms of energy consumption, but the total energy (ma- tionary SI and its past studies have proven its ability to model
chine standby energy, positioning energy, and deformation complex systems quite accurately [25, 26].
energy) consumed is the highest. This was due to the The present study proposes the feedback-based evolution-
higher speed of the forming tool. The model was formu- ary system identification (FB-ESI) approach. The proposed
lated to reduce the processing time. Similar conclusions to approach is based on the hybridization of the two SI methods
Branker et al. [15], of increasing the feed rate and step- to quantify the multiresponse characteristics of the ISF pro-
size to certain limits, resulted in significant reduction of cess. The multiresponse characteristics include the deforma-
total energy consumption. Another study [18] measuring tion energy and the unclamped geometric accuracy. The inputs
the effects of inputs (forming time, etc.) on the energy considered are step down, sheet thickness, tool diameter, and
consumption on CNC milling and CNC turning machine wall angle. The models are formed for the each characteristic,
was investigated. It was concluded that forming time has analyzed and validated, and the surface analysis performed on
the highest influence on energy consumption, and thus, by each results in quantitative relationships which can monitor
reducing the forming time by ten times, the reduction of the two characteristics simultaneously.
Fig. 4 Scatter matrix showing the nature of a deformation energy and b unclamped geometric accuracy with respect to four inputs
and the step down and the forming time. In this, the summation deformation energy (J) and unclamped geometric accuracy
of the deformation energy in three directions results in the total (mm) with higher non-linear variations. Figure 5 shows that
deformation energy. On the other hand, the geometric there is a high non-linear collinearity between the inputs and
unclamped accuracy was measured by the 3D digitizer. The the two characteristics. The fivefold cross-validation approach
3D data of the part was processed using GEOMAGIC and is used for splitting of the data into five training and testing
the dimensional errors between the formed and the CAD data datasets. The training dataset accounts for 75 % of the total
was computed. The deviation measured in the vertical direction samples while the testing set accounts for 25 % of the samples.
of the designed pyramid and targeted truncated pyramid profile
was used as global geometric error [3]. Experimental
In this way, by varying the four inputs (step down (mm), data of the
sheet thickness (mm), tool diameter (mm), and the wall angle ISF process YMODEL
YGP
GP +
(degree)), the response characteristics (deformation energy (J)
and unclamped geometric accuracy (mm)) were measured. A EGP
total of 40 samples (Table 2) were collected from the study Error component of
the GP model and the
measuring these response characteristics based on varying the experimental data
four inputs as given in Li et al. [3]. Figures 3 and 4 show the SVR
nature of two response characteristics (deformation energy (J)
and unclamped geometric accuracy (mm)) of the ISF-formed Fig. 5 Schematic diagram of the feedback-based evolutionary system
pyramid part. Figure 4 shows the non-linear nature of the identification (FB-ESI) approach
1358 Int J Adv Manuf Technol (2017) 89:13531365
Fig. 6 Bar graph with the lowest MAPE of the ESI and FB-ESI models (a) and b deformation energy for dataset 4 (c), and d unclamped geometric
accuracy for dataset 3
3 FB-ESI approach methods. GP is known for producing the explicit models and
the SVR is known for providing the generalization ability of
The FB-ESI approach comprises two potential SI methods of the models. Secondly, the notion of parallel mechanism en-
genetic programming (GP) and support vector regression sures that the predictions obtained from the GP model are
(SVR) connected in parallel mechanism with each other. In trustworthy as the SVR model connected in parallel predicts
this work, ESI is referred to as evolutionary system identifi- the error of the GP model; the summation results in the actual
cation approach of GP. The motivation behind adapting these prediction value. Figure 5 shows the parallel mechanism
two methods is to combine the positive features of the two adapted for these two methods. The SVR method takes in
the necessary input data (experimental data and predictions of model by predicting its error. To understand the notion of this
the GP model) and acts as a feedback method for the GP mechanism, each method is discussed in brief.
Y Model Y GP EGP 1
where YGP is the output value obtained from the ESI GP ap-
proach, EGP is the prediction of the error of the GP model
made by the SVR model, and YModel is the actual value pre-
dicted by the hybridized FB-ESI approach of GP and SVR.
GP [27] based on evolutionary SI follows the same mech-
anism of genetic algorithms (GAs). The only difference be-
tween them is that GP produces tree-shaped structures, i.e.,
models, whereas GAs produce crisp values for inputs that
optimize the process. In other words, GP is used for model
building processes whereas GAs are used for post model anal-
yses (optimization). The functional and terminal sets in GPs
are responsible for the determination forms of the models. The
fitness function/objective function used is structural risk
Fig. 7 Performance of the proposed models on the testing dataset: a Fig. 8 Box plot statistical distribution of the relative error of the
deformation energy and b unclamped geometric accuracy proposed models on the testing dataset
1360 Int J Adv Manuf Technol (2017) 89:13531365
minimization (SRM) [28], and the rest mechanism is the same [32]. Advantages of SVR over the traditional soft computing
as those given in past studies [29]. methods are that it is not based on any statistical assumptions
The framework of SVR is developed on the principle of sta- and is completely a data-driven black box modeling approach.
tistical learning theory of SRM [3032]. It is also referred to as The basic concept of this method is to transform the given non-
the extension of support vector machine algorithm, which was linear problem to a linear problem by projecting the input space
used to solve classification problems. The ability of SVR in into a higher dimensional space using a kernel function.
modeling complex symbolic regression problems has been found Complete details about this methodology can be found in [32].
The methods are implemented in MATLAB R2010 with the Appendix are the formulated models from the datasets. These
parameters such as population size; number of generations; depth models are analyzed statistically in Sect. 4.
of the tree; tournament size; and number of iterations set at 300,
100, 8, 7, and 10, respectively. The kernel function used for SVR Mean absolute percentage error MAPE%
is radial basis functions with its parameters (width and sigma)
1 X Ai M i
estimated using the coupled simulated annealing and grid search 100 2
n i Ai
methods. The bar graph (Fig. 6) reveals that the mean absolute
percentage error (MAPE, Eq. 2) of the models corresponding to
dataset 3 (unclamped geometric accuracy) and dataset 4 (defor- where Mi is the value predicted by a model, and Yi is the actual
mation energy) is minimum. Equations A1 to A4 in the value of the output.
5 Conclusion ness and true prediction value for the ESI model. The pro-
posed models can be used in industries to predict and monitor
The present work highlights the motivation of incorporating the two characteristics offline, thus saving some amount of
system identification theory in modeling two response char- experimental resources and time. Parametric analysis reveals
acteristics of the ISF process. In this context, the comprehen- that the sheet thickness influences deformation energy of the
sive study based on the FB-ESI approach is proposed for the ISF-formed part the most, while the wall angle influences the
quantification of deformation energy and unclamped geomet- unclamped geometric accuracy the most. Future work for au-
ric accuracy based on four inputs (step down, sheet thickness, thors is to include additional parameters such as machine type,
tool diameter, and wall angle). Statistical analysis based on the spindle speed, and final depth and investigate their effects on
cross-validation approach and statistical error metrics con- the response characteristics (deformation energy and
cludes that the proposed approach outperformed the ESI ap- unclamped geometric accuracy) of the pyramid part fabricated
proach. This is because the SVR model acts as an error pre- from ISF technology.
diction model, and thus, for each instance that the deformation
energy and geometric accuracy is predicted from the ESI mod- Acknowledgments This study was supported by the Shantou
el, the value of error adds to it, giving a sense of trustworthi- University Scientific Research Funded Project (Grant No. NTF 16002).
Appendix
ESI j 6820:6584
Deformation energy
6:4104* tanhtanhx1 x1 tanhx1*x2*x4*x2
2:672966 *sintanh6:306133 611:5618* expplogx2 x4
2:672966*15:079500*x4*x2
tanplogexpx4 sintanhx4 tanhtanhx1sintantanhsinx2
82:1989*sincostanx1*15:079500*x4*x2
A1
2397:1489*expx1*sincoscosx3x2 9:0185* x4*x2
x1*x3 *x1tanhcosx3 0:78093* x4*x2
x1*x3 *x4tanhcosx3tanhcosx3
615:5297* expplogx2 x4 2:672966*15:079500*x4*x2
tanplogexpx4 sintanh13:897179 tanhtanhx2sintan8:192150
536:0318*sincostanx4tanhcosx3
Unclamped geometric accuracy FBESI mm 3:1317 0:43795* sin tanhx1
x2*x3 plogx4 *tancosx2x2 expx1 0:17119* cos cos x4 x1
x2x4 x4 x1 0:52787*tancostancosx1x2x2 0:079698
* x2*x2 expx2 tanx2 1:492886
cosx1*x4 sinx4 * tanhtanhexpx2
cossincosx4 plog0:101505 x3
3:2196*cossinx4*tanhtanhexpx2 A4
0:34202*sinsinx1*tanhx3*tanhx2 1:784e007
* tanx2*x2 expx2 tanx2 1:492886
* sintanexpx4 x1* exp tan x4 x1 x2
x4 0:043985* x2*x1 expx2 tan 1:197751
x4 cosx1*x4 sinx4*tanhx3 * tanhtanhexpx2
costanx4 plog0:101505 x3
Coefficient of determination R2
0 12
n
X
B Ai Ai M i M i C
B C
B C
B v C
B i1
C A5
u
B uX n 2 X n 2 C
@t Ai Ai M i M i A
i1 i1
v
u N
uX
u
u jM i Ai j2
t i1
Root mean square errorRMSE A6
N
MAPE RMSE
Multiobjective errorMO A7
R2
jM i Ai j
Relative error% 100 A8
Ai
Int J Adv Manuf Technol (2017) 89:13531365 1365