Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Springback Prediction in Sheet Metal Forming Combined

Finite Element Method with Date Mining Technique1


Xu Jingjing Wang Xiumei Zhang Ping Wu Yimin
School of Mechanical Engineering, Shanghai University
ANSYS Technical Support Center in East China
Chen Nianyi Lu Wencong
School of Science, Shanghai University
Abstract
This paper presents a new method to predict the springback in sheet metal forming. First the genetic
algorithm (GA) is adopted for recognizing the material parameters. According to the even design idea, the
calculation scheme is selected and the finite element method (ANSYS/LS-DYNA 5.71) is used for
calculating the springback. Compared with the calculations and experiments, the difference between the
two results is taken as source data, a new pattern recognition method of data mining called hierarchical
optimal map recognition (HOMR) method is applied for summarizing the calculation regulation in finite
element method. In the end, the mathematics model of the springback simulation has been established.
Based on the model, the calculation error of springback can be controlled within 10% compared with the
experiments.

Introduction
The springback in sheet metal forming can be described as the change of sheet metal shape compared with
the shape of the tools after forming process [1]. Sheet metals with high strength-to-modulus ratio such as
high strength steels and aluminum alloys are particularly prone to springback, and these materials are
becoming more important in automotive industry to reduce the car weight and increase fuel efficiency.
Springback makes difficult in die design because the final part shape does not conform to the tool
geometry. In order to compensate springback, die tryout is required in current automotive die development
and construction process. Die designs and construction is one of the most time-consuming steps in new car
type developing process. Therefore, how to find an effective and reliable method for springback prediction
is very important.
Currently, finite element method (FEM) is used for calculating the springback of sheet metal in forming
process, but the results of calculation are not correct, the average error is 62%[2-7]. It can’t be directly used
for die design.
There are a lot of factors, such as nonlinear, large deformation, friction, impact and material characters,
which will influence on the sheet metal forming and springback. Because the data mining technique is a
good method to treat this complex factors and data [8-11], in order to predict the springback correctly in sheet
metal forming. A new method is developed for springback prediction. FEM is used to calculate the
springback compared the calculations with experiments, the difference between the two results is taken as
source data, a new HOMR pattern recognition method of data mining is adopted for summarizing the
calculation regulation in FEM. According to the hierarchical optimal map recognition (HOMR) method to
build the hyper-polyhedron model, the mathematics expression of the springback simulation is established.
Based on the mathematical model, the springback can be predicted and the error of springback prediction
can be controlled within 10% compared with the experiment results.

1
Sponsored by China Ford Fund
FEM Calculation Model
Sheet metal forming process is a typical impact contact topic; it refers to non-linear problems, such as large
displacement, large rotation, friction and contact [12]. From the Mechanics viewpoint the forming process is
a movement and interactive action of multi-body contact system, for the forming process, the response of
the system vs. the time has to be examined. The response can be determined by the following governing
equation [13].
1. Sheet metal forming process can be described by motion equation
..
σ ij , j + ρf i = ρ x i (1)

..
Where σ ij is Cauchy stress, f i is body force density, x i is acceleration, ρ is current density.

2. Solution of motion equation should satisfy the following traction boundary conditions.

σ ij n j = t i (t ) (2)

Where nj, j=1,2,3, is a unit outward normal to current boundary. ti, i=1,2,3 is surface force.
3. Displacement boundary conditions can be expressed as follows:

xi (ai , t ) = Di (t ) (3)

Where Di(t), i=1,2,3 is given displacement function, x i(ai ,t) is initial coordinate of particle ai at t moment.
4. The contact discontinuity conditions along the sliding contact surface area is:

(σ ij+ − σ ij− )n j = 0, xi+ = xi− Along the interior contact boundary (4)

According to the above equations and boundary conditions, formula can be obtained by the principle of
virtual work.
.. (5)
δπ = ∫ ρ x i δx i dV + ∫ σ ij δx i , j dV − ∫ ρf i δx i dV − ∫ t i δx i dS = 0
V V V SI
After discretization the following equation can be solved by finite element method
M &x&( t ) = P ( x , t ) − F ( x , x& ) (6)
Where M is the global mass matrix, &x&(t ) is the vector of global nodal acceleration, P is the vector of global
forces (including nodal forces, surface loads and body loads), F is the vector of equivalent nodal forces
composed by element stress.

Establishment of Mathemetics Model for Springback Prediction


in Sheet Metal Forming

Parameters Selection for FEM Calculation


Metal stamping process includes forming and springback; these two steps are relevant and also
independent. The stress status of two steps is different no matter in forming step or in springback step. So
the simulation technique of the two steps is also different in application of finite element method [14].
First the dynamic explicit method is used for calculating the forming, and then inputting the previous
calculations to calculate the springback, the static implicit method is adopted for calculating the springback.
According to this procedure the computation efficiency is more improved. But the final calculation results
are different from the experiment results due to inputting different parameters in FEM calculation. In order
to study the effect of inputting parameters on the FEM results, the appropriate parameters should be
selected. The following five parameters such as the virtual stamping speed v, the penalty factor fc, the ratio
of die mesh density to blank mesh density ed, the ratio of die fillet radius to blank mesh density eb, friction
coefficient f are examined.

Determining the Source Data

Determination of Material Model and FEM Calculation


Based on the stamping experiment in Fig.1, the one-fourth-experiment device is taken as the finite element
model for calculation in Fig.2.

Figure 1 - Experiment Device

Figure 2 - Finite Element Model

The stamping experiment is carried out on the ZWICK testing machine. The angle θ1,and θ2 are measured,
see in Fig. 3, θ1=71.6°, θ2=96.8°, The ANSYS/LS-DYNA 5.71 is used for simulating forming and
springback procedure.
Figure 3 - Measured Angle in Experiment

The material character constants and invariable geometrical parameters for calculation are in Table 1

Table 1 - Material Character Constants And Invariable Geometrical Parameters


Blank density Blank Blank Blank Blank Blank length Blank holder
ρ (Kg/m3) Modulus Poisson’s ratio thickness Τ width Β L (m) length
Ε (Pa) µ (m) (m) Lp (m)
7850 2.04e11 0.271 3.4e-4 2.5e-2 7.5e-2 6.4e-2
Die density Die Die Poisson’s Punch fillet Die fillet Stroke Blank hold
ρ (Kg/m3) modulus ratio radius radius S (m) Force
Ε (Pa) µ Rp (m) Rd (m) P (N)
7820 2.07e11 0.29 2e-3 5e-3 4.3e-2 38.9
The element type shell 163, the Belytschko-Leviathan algorithm and the rate sensitive power law plasticity
material model in equation (7) are used for calculating the forming and springback in sheet metal forming
[15][16]
. The material model is:

 1

 ε& 
= 1 +    kε
p
σ n
(7)
y
  c  
 

Where σy is yield stress, ε& is strain rate, ε is strain, C, P are Cowper-Symonds strain rate coefficient, k is
strength coefficient, n is hardening index.
After the tensile test of sheet metal, based on the stress-strain curves the material model coefficients in
equation (7) are identified by our genetic algorithm code in Table 2

Table 2 - Material Model Coefficients


Coefficients C P k n
Value 0.062 46.5 300 0.13

Generating Source Data


Many factors can influence on the FEM calculation results. In this paper, except above inputting constant
coefficients and invariable parameters, the five variable parameters, v, fc, ed, eb, and f, are examined and
their variable range is shown in Table 3. Here, the even design idea is adopted to arrange the FEM
calculation scheme [17]. According to the even design table u30(3013) and its available table, changing the
value range of the five parameters, v, fc, ed, eb, and f , 30 sets of data sample are taken as the FEM scheme
for calculation angle θ1* and θ2* in Table 4.

Table 3 - The Value Range Of Five Parameters


v ( m/s) fc ed eb f
0.133-13.453 0.001-0.1 0.5-5 1.25-6.25 0.08-0.2

Table 4 - 30 Sets Of Data Sample And Their FEM Calculation Results


v (m/s) fc ed eb f θ1* (°) θ2* (°)
0.133 0.011 2.52 4.18 0.179 73.36 96.45
0.593 0.025 4.69 1.94 0.154 71.6 103.1
1.052 0.039 2.05 5.04 0.130 77.25 98.1
1.511 0.052 4.22 2.80 0.105 76.775 101.6
1.971 0.066 1.59 5.91 0.080 79.15 97.3
2.43 0.080 3.76 3.66 0.183 66.36 99.46
2.889 0.093 1.12 1.42 0.159 75.35 106.3
3.349 0.001 3.29 4.53 0.134 77.55 97.15
3.808 0.015 0.66 2.28 0.109 79.5 102.5
4.267 0.028 2.83 5.39 0.084 66.85 97.5
4.726 0.042 5.00 3.15 0.188 74.95 103.5
5.186 0.056 2.36 6.25 0.163 84.25 96.5
5.645 0.069 4.53 4.01 0.138 66.13 99.8
6.104 0.083 1.90 1.77 0.113 76.9 104
6.564 0.097 4.07 4.87 0.088 57.87 101.4
7.023 0.004 1.43 2.63 0.192 74.45 101.2
7.482 0.018 3.60 5.73 0.167 76.87 98.26
7.942 0.032 0.97 3.49 0.142 70.1 101.3
8.401 0.045 3.14 1.25 0.117 67.65 107.4
8.86 0.059 0.50 4.35 0.092 90.6 100.16
9.32 0.073 2.67 2.11 0.196 72.65 104.2
9.779 0.086 4.84 5.22 0.171 75.4 98.68
10.238 0.100 2.21 2.97 0.146 73.36 100.9
10.697 0.008 4.38 6.08 0.121 93.5 99.2
11.157 0.021 1.74 3.84 0.097 75.075 101.8
11.616 0.035 3.91 1.59 0.200 68.12 108.86
12.075 0.049 1.28 4.70 0.175 74.2 104.6
12.535 0.062 3.45 2.46 0.150 67.4 107.5
12.994 0.076 0.81 5.56 0.126 81.96 99.37
13.453 0.090 2.98 3.32 0.101 71.2 106

Establishing the Mathematics Model of Springback Prediction by Using Pattern


Recognition Method
Pattern recognition means that samples in multi-dimensional space are projected on 2D space and separated
from different classes by using appropriate method such as principle component analysis method (PCA),
linear map method (LMAP), partial least square method (PLS) and Fisher vector method, etc [11]. But for
complicated data, one method cannot separate the samples from different classes completely. In this paper,
the hierarchical optimal map recognition (HOMR) method is adopted to build the mathematics model. First
the data samples are divided into good samples and bad samples according to the given conditions. HOMR
method combines with PCA, LMAP, PLS, Fisher methods to obtain the different kinds of samples and their
maps in 2D space, and then the best recognition map is selected as the first map. The first map samples are
taken as new data to form second map, after several projections, the final map is obtained. The intersection
of all maps creates the hyper polyhedron, which includes good samples, and its boundary equations are
used to build the mathematics model.
1)Evaluation of the source data
Here, θ=270°-(θ1+θ2) means the global measured angle, and θ*=270°-(θ1*+θ2*) means the global
calculation angle, and dθ=θ*-θ are described the difference between FEA calculations and experiments. It
is assumed that the error is less than 10%, that is |dθ|<10.16, dθ is defined as target function, and the
samples satisfying |dθ|<10.16 are good sample defined as class 1, the others are bad sample defined in
class 2. Using v, fc, ed, eb, f as feature variables, dθ as target value, sorting and classifying the above source
data, the new data are created in Table 5.

Table 5 - Classifying Data


NO. class dθ v fc ed eb f
15 1 9.13 6.564 0.097 4.07 4.87 0.088
10 1 4.05 4.267 0.028 2.83 5.39 0.084
6 1 2.58 2.43 0.08 3.76 3.66 0.183
13 1 2.47 5.645 0.069 4.53 4.01 0.138
1 1 -1.41 0.133 0.011 2.52 4.18 0.179
18 1 -3 7.942 0.032 0.97 3.49 0.142
22 1 -5.68 9.779 0.086 4.84 5.22 0.171
23 1 -5.86 10.238 0.1 2.21 2.97 0.146
8 1 -6.3 3.349 0.001 3.29 4.53 0.134
2 1 -6.3 0.593 0.025 4.69 1.94 0.154
28 1 -6.5 12.535 0.062 3.45 2.46 0.15
19 1 -6.65 8.401 0.045 3.14 1.25 0.117
17 1 -6.73 7.482 0.018 3.6 5.73 0.167
3 1 -6.95 1.052 0.039 2.05 5.04 0.13
16 1 -7.25 7.023 0.004 1.43 2.63 0.192
5 1 -8.05 1.971 0.066 1.59 5.91 0.08
21 1 -8.45 9.320 0.073 2.67 2.11 0.196
25 1 -8.475 11.157 0.021 1.74 3.84 0.097
26 1 -8.58 11.616 0.035 3.91 1.59 0.2
30 1 -8.8 13.453 0.09 2.98 3.32 0.101
4 1 -9.975 1.511 0.052 4.22 2.8 0.105
11 1 -10.04 4.726 0.042 5 3.15 0.188
27 2 -10.4 12.075 0.049 1.28 4.7 0.175
12 2 -12.35 5.186 0.056 2.36 6.25 0.163
14 2 -12.5 6.104 0.083 1.9 1.77 0.113
29 2 -12.93 12.994 0.076 0.81 5.56 0.126
7 2 -13.25 2.889 0.093 1.12 1.42 0.159
9 2 -13.6 3.808 0.015 0.66 2.28 0.109
20 2 -22.36 8.860 0.059 0.5 4.35 0.092
24 2 -24.3 10.697 0.008 4.38 6.08 0.121

Now evaluate the data in the Table 5. Getting the hyper polyhedron criterion division rate R1=100% and
the division rate R2=87.5% after exchanging the class. Hyper polyhedron criterion division rate is over than
80% and it is proved the data structure is good for using to build mathematics model.
2) Building mathematics model
First using different method such as fold function method, PLS regression method, and fisher index method
identifies the data type. The predictive residual error square sum (PRESS) of PLS, PRESS = 0.83, is gotten,
then after Fold Function transform calculating the PRESS defined as FOLD_PRESS, FOLD_PRESS =
0.66.Due to Fold_PRESS<PRESS,the data type is included type. For included data, after running fold
transform calculating the fisher index of feature variables, we can get the relevant importance figure of
feature variables in Figure 4. Fig.4 shows that feature variable eb and ed have more influence on target
value dθ than feature variables v.

Figure 4 - Relevant Importance Of Feature Variables

During the FEA simulating procedure,we hope to get the best match of variables, f, fc, v, ed, eb for good
target. So we append 10 random samples using as prediction but not for building mathematics model.
Now we utilize the obtained data including 30 samples as training collectivity of the pattern recognition and
use HOMR method to get the map in Figure 5.

Figure 5 - HOMR Map


Figure 5 shows that the samples of class 1 are separated absolutely, at the same time we get the
mathematics model describing the boundary equations of good samples area by HOMR method. This
model is expressed by a set of listed mathematics inequality (8) and it is showed that the calculation error
of satisfying the inequality is less than 10%.
-1.503<=-4.188E-2[v]+4.628[fc]-0.1170[ed]-0.1379[eb]-0.3698[f]<=-0.6881
0.6359<=-3.046E-4[v]+5.270[fc]-9.370E-2[ed]+9.567E-2[eb]+4.924[f]<=1.338
0.2288<=-5.089E-3[v]-1.894[fc]+0.1655[ed]-4.297E-3[eb]+1.575[f]<=1.006 (8)
-0.3318<=+3.746E-2[v]-1.423[fc]-1.492E-2[ed]-6.718E-2[eb]+1.362[f]<=0.4925
Fig.5 also shows that there are 4 random samples in good sample area, 4 random samples see in Table 6.
That is mean that if the value of parameters variables, f, v, fc, ed, eb, are chosen by table 6, the simulation
error is less than 10%. For verifying the correction of the mathematics model, 4 random samples in good
sample area are chosen to calculate by FEM. Finally the good results |dθ|<10.16 are obtained in Table.6.

Table 6 - Random Samples Of Good Sample Area And Samples Verifying

v fc ed eb f θ1* θ2* dθ
6.89278 0.0493 3.2673 3.6483 0.1379 59.6 101.2 7.6
6.066158 0.0509 3.1139 3.5095 0.1465 69.5 99.1 -0.2
6.258457 0.097 4.1432 4.7928 0.0893 59.55 101.6 7.25
6.508642 0.0968 4.0911 4.955 0.0787 63.2 101 4.2

Conclusion
1. The material parameter identification can be improved by using GA, and it can provide more precise
simulation of forming & springback calculation.
2. According to even design idea, using appropriate inputting parameters can improve the efficiency of
forming &springback calculation. Combining the pattern recognition method with FEM, the effective
mathematics model can be established and the calculation precision can be controlled effectively.
3. A new method is applied to improve the FEM calculation precision for springback prediction. First the
material model parameters in FEM are identified by genetic algorithm (GA), and the simulation
scheme is arranged according to even design idea (EDI). Then the sheet forming & springback
procedure is simulated combining the explicit method with implicit method. Then the obtained error
data between FEM calculations and experiments are analyzed by Data Mining technology. Finally the
reasonable mathematics model is built by HOMR method and the method also is verified.

Reference
1) Yu Tongxi, 1992,Plasticity Bending Theory and Application, Beijing Science Press.
2) http://www.numisheet99.univ-fcomte.fr/
3) Daw-Kwei Leu, 1997, “Simplified Approach for Evaluating Bend Ability and Springback in Plastic
Bending of Anisotropy Sheet Metals,” Journal of Materials Processing Technology, Vol. 66, pp.9-17.
4) F. Pourboghrat, K. Chung, O. Richmond, 1998, ”Hybrid Membrane/Shell Method for Rapid
Estimation of Springback in Anisotropy Sheet Metals,” Journal of Applied Mechanics, Vol.65, No.3,
pp.671-684.
5) G.Y.Li, M.J.Tan, K.M.Liew, 1999, “Spingback Analysis for Sheet Forming Processes by Explicit
Finite Element Method in Conjunction with the Orthogonal Regression Analysis,” International
Journal of Solids and Structures Vol.36, pp.4653-4668
6) Finn M J, Galbraith P C, et al, 1995,”Use of Coupled Explicit-implicit Solver for Calculating
Springback in Automotive Body Panels,” J.Mats.Processing Tech., Vol.50, pp.395~409.
7) Narkeeran Narasimhan, Michael Lovell, 1999, “Predicting Springback in Sheet Metal Forming an
Explicit to Implicit Sequential Solution Procedure,” Finite Elements in Analysis and Design, Vol.33,
pp.29-42.
8) Forcellese A, Gabrielli F, Ruffini R. 1998, “Effect of the training size on springback control by neural
networking in an air bending process,” J. Mats. Processing Technology, Vol.80-81, pp.493~500.
9) Ruffini, R. et al, 1998,”using neural network for springback minimization in a channel forming
process, “SAE Trans. J. Mater. Manufacture, Vol.107, pp.65.
10) J.C.Liu et al., 1999,”The application of neural networks in prediction of springback of L-shape band,”
Metal Forming Technology, Vol.28, No.6, pp.0-11.
11) Chen Nianyi, Sep.2000, “Pattern Recognition Optimal Technology and Application, “CHINESE
PETROCHEMISTRY Press.
12) Zhong Zhihua, 1998,”Computation Simulation and Application in Sheet Metal Forming,” Beijing
University of Science and Technology Press, pp.142-145.
13) LSDYNA3D Theory Manual, May 1999
14) Bradley N.Maker, aug.1998, User’s Guide to Static Springback Simulation Using LS-DYNA,
Livemore Software Technology Corporation.
15) LS-DYNA Keyword User’s Manual, Livemore Software Technology Corporation, 5.1999.
16) Pan Zhengjun, April 2000, “Evolution Algorithm” Qinghua University Press.
17) Fang Kaitai, May1994, “Even Design and Even Design Table,” Beijing Science Press.

You might also like