Neural and Machine Learning to the such as automobile panels) and in analysing the

Surface Defect Investigation in mechanism embedded between the geometrical

parameters and the wrinkling.
Sheet Metal Forming
Xiaodan Wu’, Jianwen Wang’, Our research aims to investigate the geometrical
Andrew Flitman’, Peter Thomson’ influence on wrinkling in sheet metal forming in
’School of Business Systems automobile manufacturing, where more than
*Departmentof Materials Engineering seventy percent of car parts are produced through
Monash University, various sheet metal forming operations. In the
Clayton, Vic. 3168, Australia first stage of the research project, the geometrical
jenny.wu@infotech.monash.edu.au
aflitman@infotech.moansh.edu.au parameters of various formed parts used in
jianwen.wang@eng.monash.edu.au automobiles were generalized. Based on that, this
peter.thomson@eng.monash.edu.au paper explores the correlation between the
geometrical parameters and the surface
Abstract behaviour of those formed parts. We have
Surface defects such as wrinkling and buckling are a utilized Artificial Neural Networks and Machine
serious quality problem in sheet metal forming Learning technique (the decision trees method) to
industry. This paper presents using information estimate the occurrence of wrinkling based on a
processing techniques (Artificial Neural Network and set of selected geometrical variables.
Machine Learning approaches) to study the
geometrical influence on the formation of wrinkling
for the automobile components.
2. The Research Problem
The research is an attempt to investigate the
1. Introduction geometrical influence on wrinkling. The study
was carried out utilizing various automobile
In the sheet metal forming operation of
automobile manufacturing, wrinkling is a kind of drawing parts processed in the TOYOTA Press
defects that damage the dimensional accuracy Shop, including:
oil panel; plate front spring support;
and acceptable aesthetics of products. Pont spring support No. 2;
Understanding the mechanism of the wrinkling pant spring support (pressed in 4 stages, with
formation so as to predict and, therefore, to depths of 1100 mm, 1080 mm, 1060 mm,
implement procedures to avoid such surface flaw and 1040 mm respectively).
of products is an important issue to the
production engineers and researchers. The experimental data for carrying out the study
of this paper was derived through a series of
Wrinkling can be affected by various factors. experimental procedures, including a priori- and
Large number of publications studied the a posterior-forming measurements and pre-
problem from the point of view of the material processing of raw data (which had strong
characteristics of the forming parts. Engineering engineering emphasis and hence will be reported
considerations, however, arise quite frequently elsewhere). To this stage, the research problem
from geometrical factors. It is an well-accepted of ‘investigating the geometrical influence on
concept that the geometrical characteristicsof the wrinkling in sheet metal forming’ can be
formed parts are also significant factors, which formulated as an estimation task: giving a set of
influence the formation of wrinkling. Duncan & explanatory variables (the geometrical factors) to
Altan [4] and Ding & Davis [3] have taken estimate the possibility of an event happening,
important steps in this research direction but that is, wrinkled or not wrinkled.
further publications are in shortage. This is due
to the difficulties in generalizing the geometrical Six variables were considered. Three of them are
parameters (especially for complex components the geometrical parameters (see Appendix):

0-7803-S871-~/99/$10.000 1999 IEEE 1088

 XI:Curvature gradient, available, a procedure known as cross-validation
X3:Torsion gradient, was employed.
XS:Height gradient.
The other three are binary variables which Cross-validation is a method of re-using sample
indicate the boundary conditions (whether there for assessing predictiodestimation risk. It makes
are any “unusual points” existing nearby) of the maximally efficient use of, and minimal
measured block regarding respectively the t h e e assumptions about, the available data [SI. This
geometrical parameters: method, as well as its modified version known as
X2:Boundary relations in curvature gradient, v-fold cross-validation has been successfully
X4:Boundary relations in torsion gradient, applied in previous studies [5,9]. It has the
X6:Boundary relations in height gradient. potential to circumvent the problems of the
simple training-test data splitting method. In
The estimated event, or the dependent variable, is
applying v-fold cross-validation, the resultant
the occurrence of wrinkling, which falls into a
model will be assessed based on a large portion
binary type of outcome: Y = 1 meaning
of the observations (eg. 87.5% for v = 8) that
“wrinkling” and Y = 0 meaning “wrinkle-free”.
should differ only slightly from the model
developed based on all of the observations.
3. The Generalization Ability of the Model
The research task is to develop estimation In this paper, we implemented the v-fold cross-
models to analyze the relationship between the validation procedure, where v = 8, to assess the
occurrence of wrinkling and the features of a set generalization ability of the estimation models
of geometrical factors. The outcomes are developed. Each validation set had 11 selected
expected to be applied back to the manufacturing observations (3 from the Y =1 group and 8 from
process as useful tools for the product design the Y = 0 group, making 88 observations in total).
engineers as well as the workshop practitioners, This also provided 8 training sets for each having
Thus, the capacity of the resultant models for 45 observations (1 1 from the Y = 1 group and 34
generalization (ie. how accurate a model will be from the Y = 0 group, 360 observations in total).
in processing future data) should be assessed.
4. Neural Network Modelling
The concept of generalization can be defined
precisely as prediction risk [9]: the expected Artificial Neural Networks ( A N N s ) are a
performance of an estimator on new biologically inspired computation technique
observations. It represents the future performance designed to emulate or mimic the functions of the
of a model. Hence, the value of the prediction human brain. They are dynamic computational
risk cannot be calculated directly, but rather must systems where, rather than relying on a number of
be assessed from the available data. The standard pre-determined assumptions, data are used to
assessment method is dividing the experimental form the model. The technique is therefore
data into two subsets for training and validation. potentially capable of handling the noisy and
The prediction risk is calculated based on the approximate data that are typical in the domain of
model performance in processing the validation- defect and failure analysis. When using it, model
subset (observations not being used in model developers do not need to deal with the
construction). assumptions which are imposed by statistics and
which limit their modelling ability. Much
For this project, the process of derivation of research has been conducted using the A N N s to
experimental data is very time consuming and solve various kinds of estimation problems [1,6].
complex. At the time of undertaking research for This paper applies ANN to an engineering
this paper, the sample data at hand was 56 problem: estimation of the wrinkling in the sheet
observations. Given the limited amount of data metal forming operation of car manufacturing.

I089
4.1. Model development 0 A feedforward BP network consisting of multiple
hidden slabs (2 and 3 respectively) with different
In a neural network, the calculations performed activation functions.
at each neuron are determined by an activation 0 Probabilistic neural networks.
function which may be of various mathematical
form such as logistic, linear threshold, or hard 4.2. Network training and validation
limiting (odofl). The intensity of the signal
passed between any two neurons depends on As mentioned before, a v-fold cross-validation
both the activation function and the weight of the re-sampling method was implemented for this
connection. While connection weights will be study. As a result, eight training sets and eight
modified during training of the network as validation sets were obtained and used for
observation patterns are passed along, activation conducting the experiment. Each of the training
functions should be decided before the network sets was %&er split into a training subset and
training. The selection of an activation function test subset. This allowed us to modify andor to
for the hidden layer is most important, since this stop the network training based on the model
is the layer that actually performs the feature performance in processing “quasi-new data”,
extraction from the patterns processed. thus, ensuring that the resultant network model
Accordingly, we experimented with different would have better ability for generalization.
forms of activation functions, and found that for
the given problem the best suited combination of The network is trained using basic supervised
the activation functions are as follow: learning error-backpropagation. The algorithm
first layer: linear function 1-1, 11; works by presenting each training vector (inputs-
hidden layer: Logistic function; output variable pair) in turn, computing the
output layer: Logistic function. estimation error, and thereby determining
interconnection weight changes through the
Selection of the number of hidden neurons is network. The observations were presented to the
another basic decision to be made in building a network in random order in the experiment to
neural network. Heuristics have been suggested, minimize bias due to the network memorizing
but care must be taken when using them in the position of the training data. After every 180
practice as the issue is case dependent and an epochs (a complete pass of the training data
inappropriate decision may degrade the accuracy through the network), the partially trained
of the networks in generalization. At current network was tested using the test subsets
stage of ANN technology development, such described above, and the average error in
selection is still more of an art than a science. We estimating the occurrence of wrinkling was
therefore tested a wide range of hidden neuron recorded. If there was an improvement in the
numbers in order to obtain an optimal model. estimation accuacy, the network parameters were
saved. Thus, the performance of the network in
With regard to the interconnections between its processing of test data rather than of training
neurons, we considered the following network data was used to determine the quality of the
architectures: network. This approach helps to reduce the risk
0 A feedforward BP network with standard of saving a network which has memorized the
connections, ie, neurons in each layer connected features of training data, without the ability to
only to neurons in the immediately previous layer; generalize on new data.
1,2 and 3-hidden layer structures were tested.
0 A feedfoward BP network with jumped After the training session was completed,
connections,ie. neurons in each layer connected to validation of the model was performed using the
neurons in the every previous layer (jump validation data set (observations that were not
connection); 1, 2 and 3-hidden layer structures exposed to the network during its training). The
were tested.

I090
outputs of the model in processing this data set, 6. Results and Discussion
compared with the target values, represent a kind
of objective assessment of the model quality. The outputs of the neural network and the
decision tree models can be expressed as the total
5. BRP Approach Modelling percentage of the wrinkledwrinkle-free blocks
correctly predicted, and more crucially, the total
Binary Recursive Partitioning (BRP) is a branch percentage of wrinkled blocks correctly
of the Machine Learning methodology. The B W estimated. The usefulness of a prediction or
technique has its distinguish advantages in that estimation model should also be assessed from
the binary tree-structured analysis outcomes offer other points of view, such as the error stability of
explicit explanation of the evaluation results for the model and its ability to generalize. In order to
the problem under discussion. make comparisons and to evaluate the overall
performance of each model as a tool for
In this study, the algorithm used to implement estimating the occurrence of wrinkling, a group
the decision trees is the Classification and of six statistics were chosen as standard
Regression Trees [2]. A common technique assessment criteria. They are:
among the first generation of the BRP algorithms % Correctly Estimated, which measures the
was to continue splitting (or growing the tree) percentage of the observationscorrectly estimated;
until some goodness-of-fit criterion failed to met, M E , Mean absolute error, which measures the
and that often yielded erroneous results. To avoid average error rate with the best being nearest to 0;
the problem, CART proceeds by means of a STDEV, standard deviation , which measures the
three-step implementation algorithm. Firstly it dispersion of the errors that a estimation model
grows the classification trees until firther generates around the mean value of these errors;
growing is impossible. Secondly it prunes away the best is that nearest to 0;
branches of the maximal tree to get a set of sub- Misclassification Rates, categorized as Type I
trees. Thirdly it tests the error rate or costs for all Error (the frequency of incorrect estimations of
wrinkled blocks as wrinkle-free) and Type I1Error
the possible sub-trees and the maximal tree to
(the frequency of incorrect estimations of wrinkle-
determine the best tree. Based on this
free blocks as wrinkled);
combination of exhaustive searching and
RSQ, the square of Pearson product moment
computer-intensive testing, no valuable correlation coefficient, which measures the
information would be lost. correlation between the estimated and actual
distributions, ranging from 0 @oar fitting to the
Figure 1 illustrates the estimation model model) to 1 (good fitting);
constructed for the problem under consideration. Cross- Validation Prediction Risk, which measures
the iiccuracy of a model when being applied to
-9.701 N,,& 1 new data (i.e. the model's generalization ability),
I .%3 1 the small, the better.

The performance of the resultant models in

processing the training data and validation data is
given in Table 1. It can be seen that according to
the given geometrical features of the analyzed
surface areas, both of the ANN and the BRP
models can correctly classify a large percentage
of the occurrence of wrinkling (with the ANN
model performing marginally better than the
BPR model). Both the ANN and the BRP models
Fig. 1:The Optimal Classification Tree have reasonably low error rate (with the BRP

1091
model having the lower value of MAE) and that took the second and the third places in
stable error structure (small STDEV value). making contribution to the final estimation
Inspecting the model performance on processing results. The Curvature Gradient is the variable of
the training and the validation data, no significant the least significance. For the BRP model, it is
deterioration is traced which indicates that these noticed that the surface geometry parameter
models have fhirly good ability to generalize. In Torsion Gradient is the variable that contributed
other words, the models suit for future most (of the largest value of the Variable
applications. The low value of the ‘Prediction Importance). The Height Gradient and the
Risk’ observed also supports this conclusion. Boundary Condition of Height are the variables
that played the second and the third important
It is noted that the misclassification rates of the roles. The Curvature Gradient is again the
Type I Errors produced by the resultant models variable of the least contribution.
were quite large. In case of the ANN model, the
rates were around 36.36% for training data and According to the above, we can then infer that
33.33% for validation data. In case of the BRP surface geometry parameters Torsion Gradient
model, the rates were zero for training but and Height Gradient are the most important
20.00% for validation. This is the unsatisfactory factors in regard to the possible formation of
aspect of the analysis results, to which further wrinkling at a given area. The boundary
study will be directed. conditions of the torsion gradient, the height
gradient and the curvature gradient are the
Besides to examine the prediction results, it is parameters that each plays a role of certain level
also valuable to consider the relative importance of importance. The surface geometry parameter
of the input variables in producing the estimation Curvature Gradient, however, seems not
outcomes for each model. For a neural network, a influencing much on the process of wrinkling
parameter known as the Contribution Factors formation. In spite of being preliminary, these
(CFs) can be observed. These factors are the sum are very useful findings, given the fact that there
of the absolute values of the network connection is no established theory available on describing
weights, leading from each neuron that the mechanism of the (surface defect) wrinkling
represents an input variable. In this regard, the Occurrence and the surface geometry conditions.
CF value of a variable gives an approximate In light of these, new experiments will be
measure of the importance of that variable designed and carried out in our follow-up
relative to the other variables in the same research. That will hrther examine and test those
network. By examining the CFs, some useful results, aiming at more clear understanding of the
information with respect to the relations between mechanism of the geometrical influence on the
input variables and the estimated target variable wrinkling formation.
may be obtained. Similarly, for each input
variable of a BRP model the Variable 7. Conclusions
Importance can be used to observe that variable’s
contribution in deriving the final estimation. This paper used the neural and the machine
learning techniques to estimate the occurrence of
Table 2 lists the CFs and the VariabZe wrinkling of the sheet metal forming parts of
Importance of the input variables corresponding automobile manufacturing. A neural network and
to the surface geometrical parameters of the a binary recursive partitioning classification tree
estimation models. For the neural network, the model were developed which demonstrated fairly
surface geometry parameter Height Gradient is good performance in processing the training and
the variable that contributed most (of the largest the validation data. The input explanatory
CF value). The Torsion Gradient and the variables of the models were a set of geometrical
Roiindary Condihori of Torsion are the variables parameters of the forming parts. Thus certain

I092
kind of relationships between the surface References
geometry features and the wrinkling Occurrence
1. Altman, E,,Marco, G., and Varetto, F., Corporate
of the corresponding area of the parts have been distress diagnosis: comparison using linear
established. The estimation models would be discriminant analysis and neural networks (the
useful tools for the product design engineer as Italian experience), Journal of Banking and
well as the workshop practitioner in their Finance, vol: 18,pp. 509-525,(1994).
tackling the problem of surface defects of the 2. Breiman, L.,Friedman, J., Olshen, R. and Stone,
sheet metal forming parts. C., Classification and Regression Trees,
Wadsworth, Pacific Grove, USA, (1984).
3. Ding Q. & Davis B.J., Surface Engineering
Appendix: The parameters for Surface Geometry Geometry for Computer-aided Design and
1. Curvature is the rate of change of the direction of Manufacture, Halsted Press, England, (1987).
the tangent vector of the curve with respect to arc 4. Duncan J. L.,Altan T., New Directions in Sheet
length: K(S) = Im(s)/dsl. Where K is curvature, T is Metal Forming Research, CIRP Annals, vol: 29,No.
tangent vector, and s is arc length. If the curve is 1, pp, 153-156,(1980).
described by the arc length parameter equation 5. Fletcher, D. and Goss, E., Forecasting with Neural
R(s) = IX@) Y(S) ml Networks: an application using bankruptcy data,
then K(s)=/r ~ s ~ ~ = [ x ( s ~ ~ + ~ ( s ~ ' + z ~ s ~ ~ Information
~~~ &Management, vol: 24,p.159,(1993).
6. Flitman, A., Towards Analysing Student Failures:
2. Torsion is the rate of turn of the direction of the neural networks compared with regression analysis
binomial vector of the curve with respect to arc and multiple discriminant analysis, Computers Ops.
length: t(s)=ldB(s)/dsl . If the curve is described by Res. Vol: 24,pp. 367, (1997).
the parametric equation for arc length, then
7.Steinberg, D.and Colla, P., CART: Tree-Structured
z(s)= r '"0( r 'x r r "12 Non-Parametric Data Analysis, Salford Systems,
3. Height is the vertical displacement of each point CA, USA, (1995).
from its initial position: h(s)=z(s) 8. Stone. M., Cross-Validation Choice and
A program was written to calculate these three Assessment of Statistical Predictions, Roy. Stat.
parameters. It was assumed in the present work that SOC.,B36, pp. 11 1-147,(1974).
the change of the above parameters at each grid point 9. Utans,J. and Moody, J., Selecting Neural Network
is important,rather than the value itself at that point. Architecturesvia the Prediction Risk: Applicationto
The actual values to be used as the input variables of Corporate Bond Rating Prediction, Proceedings of
the neural network were the gradients of the three 1''. Inter. Con$ on Artijcial Intelligence Application
parameters at each point. on Wall Street, CA, pp. 35-41,(1991).

Results on training data Results on validation data

- ANN model BRP model ANN model BRP model
correctly Predicted 90.91% 88.64% 86.36% 83.33%
NRM!jE 0.723 0.836 0.776 0.856
MAE 0.203 0.111 0.237 0.167
STDEV 0.304 0.321 0.351 0.426
Type I Error 34.09% 0 37.50% 20.00%
TypeIIError 3.03% 14.29% 3.13% 14.29%
RSQ 0.463 0.551 0.372 0.432
Prediction Risk 0.114 0.114 0.125 0.167

Variable Geometrical Meaning

Name ANN model BRP model
x5 Height gradlent0.3685 (1) 0.2131 (2)
x4 Boundary of torsion gradient
0.1823 (2) 0.1030 (4)
x3 Torsion w ent 0.1635 (3) 0.3943 (1)
x2Boundary of curvature gradient 0.1217 (4) 0.1030 (4)
x6 Boundary of height gradlent 0.0973 (5) 0.1867 (3)
Xl( 0.0667 6 0.0005 6

I093