OPTIMIZATION OF MULTIPLE PANEL FITTING IN AUTOMOBILE ASSEMBLY
D. Knorzad, J. Shi, S.J. Hu, J. Ni
Department of Mechanical Engineering
‘The University of Michigan
‘Ann Arbor, Michigan
E. Zussman, G. Seliger
Department of Assembly Technology
“Technical University of Berin-WE
Bertin, Germany
ABSTRACT
‘Asystemalic approch is prescated to obi improved panel ft
quality through the use ofan optimum pane fing strategy. The
objective of the optimal pane! Sting strategy isto determine the
location of the panels on the automobile body such thatthe ap
and fhsh variation ofthe panel tae minimized, This approach
‘ses measurement data from both the panels and the bedy-in-
‘white (BIW) to determine the optimum position of mukiple
panels im an automobile Body opening. Fist, some indices are
efned to quantify the quality of «panel ft Second, the sources
of variation inthe pp nd tah are preseted, Thea te mule
panel fing problem is formlsted into contained
cptimization model, The eects ofthe optimization process for
tiple panels are then demonstrated and validated through the
se of computer simulation The computer simula
demonstrates the optimization model and algorithm by reducing
the within-car gop and Dush variation average by 24.3% and
by as mach at 45.4% inthe cae study presented.
INTRODUCTION
In today's highly competitive automobile market the quality of
the product is becoming increasingly important. Among the
‘quality concems of the automobile manufacturers i the quality of
the fit and finish. Specifically, the quality ofthe panel fits ofthe
automobile ranks high in concern due tothe high warranty cost.
‘An inadequate panel fit will contribute not only to fimctioaal
problems such as water leakage aod wind noise but secthetic
problems such as meven gap and fush.
‘The indices that ae ueed to determine the quality ofthe panel St
are the gap and Ouch variation, gap range and fush range. These
‘terms will all be defined more rigorously later in the paper. The
dimensional variations ofthe gap und flush between the BIW and
the panel, or between panels, aise ffom four sources which are
‘Transactions of NAMRUSME,
241
gue 1; (1) dimensiousl vasistion of the panels; (2)
dimensional variation of the BIW; (3) variation of the panel
Sitting process; and (4) effects of painting and general assembly
Each branch of the fishbone diagram shown bellow bas various
sources of variation which effect the gap and fush variation.
inlure Repestobity
Surface Contour
Doge Door Fil Progess
De Opering,-*™ Other Processes
rap Ate
ié
General Assembly
FIGURE 1: FISH BONE CHART OF VARIATION
‘COMPONENTS IN PANEL FITTING
‘he fest two vacations ae subassembly vavstions sod ned to
‘be improved atthe subsssembly level. These ius have been
‘sddressd by Wu and Fu (1990), The fourth source of variation
iss topic of esearch ouside the range of this paper. This paper
focuses onthe variation induced bythe pane! fiting process. In
‘his paper, a special example of panel fiting, automebile doors, is
‘wed to demonstrate the optimization process
‘The control of dimensional devistion i the stomobile dustry
has been msinly done by feeding Statistical Process Control
(PC) date back to manuactring proceses suchas samping or
par subassembly, Wa and Hu (1990), Fiting of rgd bodies in
sacral was suggested by Beans and Meng (1984) which
presents + grcral constrined optimization process of single
ody. An approach to integrate the data dizety to the Sting
Volume XXIII, 1995process was proposed by Wu, Hu, and Wu (1954) where the bes
fiting parameters are transformed 10 » door fixture location,
However, Wu etal, (1994) described the best fiting process in
‘which only one door wat ft at a time, Recently, Schule,
Zussman, and Seliger (1994) and Sakai and Yarumatsy (1993)
presented
determine the position of
location criterion, However, litle literature exists in the area of
sultple body iting optimization.
In recent years, the implementation of the in-line Optical
‘Coordinate Measurement Machine (OCMM) in the automotive
industry and the increasing accuracy of industrial robots provide
ew opportunities for the development of « flexible assembly
systems for the fiting of automobile panels, This flexible
sssembly system atlenpte to find the optimum position of the
doors inthe BIW such that the within-car gap and flush variation
is minimized. In this approach, multiple doors are to be ft
simultaneously 40 thet the dimensional devistion in the various
parts can be distributed evenly over the entire door fit of the
automobile by the optimization process, However, efictive
implementation of this multiple pane! ft optimization remains a
challenge. Some of these challenges can be summarized as
follows:
1. Transformation ofthe discrete OCMM measurement data to a
‘geometric model of the pane fi.
Formulation ofan optimizarion objective funtion to minimize
the within-car gap and flush variation.
Formdation of constraints to restrict the optimization
problem.
In this paper, «systematic approuch to obtain optimum multiple
‘panel fitting is preseated based on 2 constrained optimization
problem, The approach is demonstrated by locsting the font and
rear doors of en automobile timuianeouely. Constraints that are
taken into account are Enler parameter constraints, minimum gxp
constrsints and minimum parallelism constaints. The
‘organization of the paper is as follows: Section 2 describes the
‘rometric model of muhiple panel fitting. Section 3 presents the
constrained multiple panel optimization scheme Section 4
‘verifies the proposed method through computer simulations. In
section 5, sensitivity analysis was performed on the optimization
constraints. The conclusions of the paper are discussed in the
section.
2)
3.
GEOMETRIC MODEL
The objective of the optimization model is to minimize the
variation in the gap and Duc Essentially the pp and Dash are
projections ofthe vector, 7, from the Door tothe BIW st the
measurement points. The projection ofthis vector, 7, oato the
vectors GP; and FP; at the specific poist will result ia the
tmearure of the gp and fluc repectively. The vectors GP; and
FP; are mit vectors which define the direction ofthe pp and
‘fash at + specified point i (igure 2). The diction of these
vectors is govered by the design of the gap and fh
Essetily, the pp veto, GP, i the unt vector tangent to the
design direction ofthe vector 7, Similarly, the Buh vector, FP,
Transactions of NAMRUSME
242
is the nit vector perpendicular to the design detion of the
vector 7, An llustation of the ap and Ssh measures i showa
infgue?
i
|
actual
Door Surface
pone “I
FIGURE 2; ILLUSTRATION OF GAP AND FLUSH
Body Surtoce
‘As mentioned, + panel is defined as any sheet metal part
attached (0 the BIW. The doors, hood, fender, and butch are
examples of panels, Definitions of other terms used throughout
this paper are:
B= the coordinates ofthe th point on the body
[Dl =the coordinates of the jt point onthe freat door
Bd = the coordinates ofthe th pont onthe rear door
‘TIQK) = Front door homogeneous transformation matrix (4x4).
‘T2QK))= Rear door homogencous transformation matix (4x4)
ies Poe
rouse
Oy * Coordinate Rotational Maroc
ard see om tee
Quam Mente S+4-k 9-48
eemae ten ard-X!
‘wher, ate euler parameters
Poxt= Coordinate translation matrix=[R, R, Ry]
Whee RyRy tm ton a he) 2
coordinates. Tau,
rE K amma sete Bi/?.
enemy aed“ K onoan Bl?
sama ota ate-K Rel?
° ° oo 4
Xp th optinization variable,
{1 1-7 forthe fot dor; i~8-14 forthe rear door
To(K)= 244
OPTIMIZATION MODEL,
‘bijective Function;
Figure 3 shows some posible measurement locations for the
sutomobile door fit, The measurement points can be clastifed
into three clases. The fist class of points (Class 1) are points
whose gap md Ouch values are determined by the relative
location of the from door ond the BIW (points 1 8). The cond
class of points (Clase 1) are points whose gap and values are
‘Volume XXII, 1995Aetermined by the relative location ofthe rear door and the BIW
(points 9- 16), The third class of points (Class I) are points
‘hots gap and fish are determined by the relative location of the
vert and the rear door (pots \7 snd 18).
na mil me, Un ae
FIGURE 3: LOCATION OF MEASUREMENT POINTS.
‘The vectors fom the menrrement poiat oo the BIW to the
corresponding messurement pinto the dao ae:
3,= 7x}, a
oy
ae
1B, T(x,)eDa,
nlx,}eo1,~Tax)03,
where, j= the measurement poot number
i 123,14 (he variables tobe optimized)
Equations (e-e) ate the vectors between th door end the body
seach measurement pia fc: the door has ben placed inte the
door opening, The form of equation 1 depends on the (pe of
Print that is of interest. For example, ifthe pat of teres isa
Clas pint (1-4), the proper eqution would be eqution
1a. On the other band, if the pont of interes is» Cat I point
(914916) the proper equation would be equation Ib. Silly,
ifthe poiat of interest is «Class I pot (pt # 17 & 11), the
Proper equation is equation Le
To get struc representation of the ror i the ft, we need to
ojet these vectors cto the vectors which define the xp and
huh (figure 2). The vector that dens the direction of ap at
Print i is called GP; Similarly, the vector thet defines the
direction ofthe shat peat iis called FP, Therefore, the gap
can be defined by the length of his projected vector
Symbolic his writen forthe gpa
‘ra(2;—n1x)}04,07,)
(0;~Tlx,}02, 0%)
1i(n(x,)01,-12(x,}02, 0%)
Similarly, forthe ah a:
ra(3,—1(4 +01, )
*r9(8,~T2(x,)+D2,,F7)) @)
oi( x}, -12(x, Da, 9) ry
Where, “pro(ABY" is «function which calculates the length of
Projection the vector A oat B.
The pp and fash values that are calculated by using equations
2and3 yield the absolute pap and Huh values These vales of
1p and Dish ae the function of the optimization viable %)),
@
a
ay
cD)
‘Transactions of NAMRIU/SME
243
that sto determine the poston ofthe doors inthe BIW. To pt
the dviton from the design vale a each measurement point
these values of pap and fu canbe subtracted fom the design
(0G) and fhsh (DF;). If the deviation ofthe gap and fush
from nominal are squared and summed forall the measurement
Pint the optimization tndel objective intone
ot €)-T10K) * 94.08 BG
8 = TLR) DH FRI OR
al c= T2 1x) *02,.08,)-Da,Ps},
Hae nd (87204) 03 -F,- OF
|W x (proj (71 (X,) * Dy ~ THK) * D2. GR)~ ral
Fem es cK) * DK TK) * DAH) BA
where, I= [to No, of Clas Pots
4j=No of hss Points to No.of Clase [1 Points
‘= No.of Cas & I eit o Teal No, of Points
DG; = Desi Gap at pont
DF = Design uch t point
vers. of
Constraints
‘The constraints of the optimization problem are derived from
‘the geometry of the problem. There are three classes of
constraints and within each constraint class there are multiple
constraints. The three classes of constraints are euler parameter
‘sormalization constraints, minimum gap constraints, and
‘minimum parallelism constraints
Euler Parameter Normalizalion Constraint. The Euler
parameter constrain is inchued a » constraint to insure that the
fetation of the door is valid, therefore, resuhing in a
‘homopencou transformation, The relationship between the Euler
Parameter must be:
dedededet
1 the optimization routine however, this comraat i writen as
an inequality constraint to account for computer round-off errs.
ence, the constr fr the foot doo eatin ie writen a:
te OF ex} exX}oxL-DsE
“This constrain must also be inched for the rear dor to esse
thatthe rotation ofthe rear door is valid. Simlary, the Euler
parameter constraint forthe ee door x
or OG 9d Xho,
‘wher, Eis a small number and the Xi this case re the Euler
paramters and optimization variables (X; ~ Xj forthe front door
nd Xq =X; forthe rea door)
‘Gap Constraint. The gxp constrain is included inthe model to
allow the user some flexibility. The user has a choice in the
luge of the pap constraint. Esseataly, the gap constraint is
‘Added to the model to allow the user to specify « minimum
Allowable gap at « measurement point
‘To develop the equation for the constraint, the vector from a
‘meaturemeat point on the body to the corresponding pot on the
door must be defined. This vector has been defined in equation |
Volume XXII, 1995remant Point
usd Fl alte Femi A |
‘
T= Ate Optimization
FIGURE 4: GAP AND FLUSH VALUES BEFORE AND
AFTER OPTIMIZATION
TABLE 1: STATISTICAL RESULT OF THE CASE STUDY
Gap & Flush | Gap Range | Flush Range
Variation
‘WiO Opim. | 2.5952, 15989 23741
‘With Optim | 1.4681 0.6972, 1305
* All wits ia me
Ia the previous example the weighting factors were all st
‘quily, That is, the importance of each gap and Ouch in he
‘optimization was equal Also, there were no minimum gap and
‘minimon partis coustrins wed inthe optimization.
To father demonsiate the cxpsbiltis ofthe opinizton
scheme, te same CAD gencratd data was wod for encating 50
replctions of the optimization described above. For cach
replication the random deviation on the order witnessed in
suomotile plants (60 = 2 mm) is added to each poi. With
‘ach replication the optimum seltion is found andthe within car
‘arition before and afer optimization is calculate ad pleted in
figure 5.
For comparisoe, the plot of the before and afer opinizton
‘within ca gp od flush version for the 50 samples i shown in
figare 5, From the plot enc can soe that in evry cae the within
car pp a ach vtiaton wn redcod. On sverige he vation
was reduced by 0.69 mum. A summary of some of the
‘optimization parame of interest are shown in table 2
‘As can be Seen in figure $ and table 2, the variation and the
‘ages lave been reduced. ‘This reduction wil eu in door ft
that is more aesthetically pleasing to the customer as well as
increasing the functionality f the door.
‘Transactions of NAMRUSME 245
FIGURE &: WITHIN CAR VARIATION BEFORE AND
AFTER OPTIMIZATION
‘TABLE 2: SUMMARY OF THE OPTIMIZATION
‘SIMULATION
Reduction in
Gap and Flush 63 069 mm
Gap 66 0.59 om
“Gap Range 041 om
Flush 60, 0.69 mm
Flush Range 045 mm,
SENSITIVITY ANALYSIS:
In the model description above, two clasts of constraints are
introduced. These constraints (Minimum Gap and Minimum
Panlelism) are used only when the user chooses to use them.
‘This raises the question of how these constraints affect the
‘optimization. To determine thit enswer analytically would be
‘very dificult due tothe ooa linearity and the shear size of the
problem. As « consequence, we have choten to do a sensitivity
analysic of two classes of constraints using an experimental
‘approach. In each sensitivity analysis conducted the data from
‘the case study is used. The optimization is then conducted with
(he appropriate constraint ured. The constraint value is then
varied over s range and the optimum solution is found and
compared to the optimum solution without axy coutrints.
In the case study presented previously, the optimum location of
the panels yielded a gap and fush variation of 146 mm. Itcan be
sem from Ggure 6 thal, when 4 gap constraint is used, the
‘optioum gop and Such variation varies from the unconstrained
‘optimum value of 1.46 mum to over 4 mm. When the minimum
‘Bp constraint value is relatively mall the optianum gap and flush
variation is found to be the same as the case when no gp
constraint is used. However, when the value of the minimum gap
is increased, the optimum value of the gep and fush increases
accordingly. Otherwise sated, asthe minimum gap constraint
value increases, the minimum gap constraint becomes active.
Volume XXIt1, 1995waasaevgagagaass:
| Sean tates Sen
FIGURE 6: GAP AND FLUSH VARIATION VS. MIN. GAP
VALUE
‘Similarly, figure 7, when » minimum parallelism constraint is
‘used, the optimum gap and flush varttion varies from the
‘unconstrained optimum value of 1.46 mm to over 1.$ mm. When
‘the minimum paralleliem constraint value is relatively small the
‘optimum gap snd fush variation is found tobe the larger than the
cave when no gap constraint is used. However, when the value of |
‘the minimum gap is increased, the optimum value ofthe gap and
‘Such decreases othe unconstrained value of 1.46 mm.
etna vate Menten ae A
FIGURE 7: GAP AND FLUSH VARIATION VS. MINIMUM
PARALLELISM VALUE
‘CONCLUSIONS
The pane! fiting problem was formulsted into a constrained
‘optimization problem. The objective of the optimization is to
determine the postion of automobile panels in the body opening
such that the within-car variation of the gap and flush is
‘minimized, The constraints in the optimization model include:
() Euler parameter constraints; (2) minimum gep constraints,
snd (3) minimum parlleliem constraints, Simulations were
conducted to evalute the effect of the optimal Siting on the door
‘Gt. Based on these simulations, the following observations can
tbe made:
|The gap and fluch varistion is reduced by the optimization.
Based on the 50 simulated door ‘ts presented, the
optimization algorithm reduced the within-cervaristion
Transactions of NAMRI/SME 246
‘the gap and Dush on average by 24.3% in $0 simulated door
fits and by as much as 43.43% inthe case study presented in
this paper. Asa result, the door ft quality is improved both
sestetically and functionally.
From the $0 simulations presented, it can be seen tht the
‘optimization algorithm found the optimal postion that
reeuced the variation in every case, which illustrates the
‘robustness ofthe optimization algorithm,
REFERENCES
1
10.
u.
2
B.
Bons, B., and Mengs, G., 1984, "Best Fiting of Three
Dimensions! Bodies", IFAC 9 th International World
Congress, Budapest, Hungary
Han, SP, 1977, "A globally converge variable method
for nonlinear programming’, Journal of Optimization Theory
‘and Applicaions. Vol. 22, pp. 297-309
Hock, W., Schitkowski, K, 1981, "Test Examples for
[Nonlinear Programming Codes", Lecture notes in Economics
‘and Mathematical Systems, Vol. 187, Springer
Hoska, DR, (1988), "FLAM: What itis. How to achieve
itt, Manufacturing Engineering, April 1988, pp. 49-54
Lawsoa CL. and Henson RJ, Solving Least Squares
Problems, Prentice-Hall, Inc., Englewood Cliffs, NJ. 1974
Sakai, H,, and Yasumetsu,S., 1993, “Development of Body
Fiting System", Proceedings. of the 26 th International
‘Symposium on Automotive Technology and Automation
(ASATA), pp. 69-76, Aachen, Germany.
‘Schitkowski, K., 1993, "Mathematical Optimization - An
Introduction”, Internartonal Series of Numerical
Mathematics, Vol. 110. pp. 33-42
Schitthowaki, K, 1984, "NLPQL - An Optimization
‘Algorithm for Nonlinear Quadcatic Optimization ”
Schitthowaki, K., 1986, "DFNLP - A Subroutine to Aid
NLPQL in the Optimization of Data Fitting Problems",
Version 12
Schitthoweki, K., 1984, "Design Implementation, and tet of
4 nonlinear Programming Algorithm”, Report, institut Fuer
informatik, Universitaet Stuttgart, 1984
Schuler, H, Zussman, E., and Seliger, G., 1994, “Optimal
Fiting of Objects Based on Range Images - A Planning
‘Approach Assisted by Local and Global Measurements
Matching’, Proceadings. of the First SM. Wu Sympostum,
Vol. 1, p. 305-312, Chicago.
Wu, SK, Ha, SJ, Wu, SM, 1994 Optimal Door Fitting
‘with Systematic Fixture Adjusting In. Journal of Flexible
Mamafacturing Systems.
Wu, SM, and Hu, $J., 1990, “Impact of 100% Ia-Process
Measurement on Statistical Process Control in Automobile
Body Assembly," Monitoring and Control in
‘Mansfacturing, E4. S. Liang, T.C. Tst0, ASME Winter
‘Annual Meeting, Nov., pp. 433-448.
‘Volume XXIII, 1995