Professional Documents
Culture Documents
10.1016@0010 44859390003 7 PDF
10.1016@0010 44859390003 7 PDF
optimization techniques
u s i n g FANGA c u r v e s a s t h e
qual ity criterion
Gunnar Liden and Stefan K E Westberg
The work has been carried out at Saab Aircraft bPINE (U~
Division, which hereinafter is referred to as Saab. Li
L~
Finding a good and usable objective function is the same / /
process as studying different surface-analysis methods. / /
These methods may be divided into three groups, as // L3
Lq
follows.
/[03
• Position criteria." The normal way of checking a surface /
today is still with planar intersect curves. Intersects are
easy to interpret, and the shape of the curves often
indicates how the input data to the surface should be
1t111/
changed to improve the surface quality. It may,
however, be quite laborious to determine appropriate
intersection planes and the distance between these
J
planes. / /
/
• lst-order-derivative criteria." The analysis of reflection
lines falls into this category (see Figure 1). Smooth
reflection lines are a very good indicator of high surface
quality. However, the reflection lines are relatively Figure 2A. Conic-lofting surjace
[L: longitudinal-limit curves. TD: tangent-definition curves.]
difficult to compute, and they are of little help in the
surface-definition work, since they normally give no
hint of how the input data to the surface should be L1
modified.
• 2nd-order-derivative criteria: One example is the
Gaussian curvature presented in a coloured plot which
can be generated by many CAD/CAM systems. This plot
is very difficult to interpret geometrically, and is // _)_L2
normally of no help at all in the surface-definition /
work. /
• Step 1." Compute a point and derivatives for the given FORMELFI
U on the spine curve. The plane for the conic section
passes through the point, and the normal to the plane
is the tangent to the spine curve.
• Step 2: Make intersects with the longitudinal-limit
curves and tangent-definition curves, and compute
derivatives with respect to U (the intersecting spine
plane).
• Step 3: Compute a P value, i.e. the fullness of the
curve according to Figure 2B, and the derivatives of
P with respect to U.
• Step 4: Create a conic defined by two limit points,
two tangent-definition points, and a P value, and Z AXIS
compute the point coordinates for the given V value
and the tangent, i.e. the derivative with respect to V.
• Step 5: Compute the surface normal in this point using J
the above computed derivatives with respect to U (the
intersecting spine plane) and V. V5
iNPUT
FRNGR
VIEN XY
VJEN XZ
VIEW ISO
OUTPUT
FRNGA
Figure 5. Many surface strips can be generated from one set of FANGA curves
However, at the current stage of development, an adjustments in the surface input data that are necessary
optimization-based surface-definition process should for a high-nuality surface.
interact intimately with the loftsman's geometry The traditional input data to a conic-lofting modeller
knowledge and experience to be efficient. are points and vectors, which define the longitudinal-limit
The definition of a preliminary surface close to the and tangent-definition curves, the spine, and the scalar
final production surface can be developed using functions. The appropriate additional input data for a
traditional methods. The time that it takes a loftsman to loftsman are, for instance, a vector which defines the
define such a surface is usually not a problem, and the direction in which a curve point may be moved, and a
optimization computations are better conditioned if vector which defines how a curve tangent is allowed to
started from such a surface. rotate (see Figure 9). These geometrical-variation
It is in the last phase of the surface-definition process restrictions can be defined using standard functions in a
that the use of optimization techniques is most profikble. geometry system.
It is extremely time-consuming to make the small We believe that the most efficient way to define a
STRIP i
5TRI? i STRTP 3
S~RXP 2 Sq-R~P Ll
surface is by (sub-)optimizing parts of the surface, i.e. compromise, a result of discussions between departments,
letting only a few points and vectors (up to 20), which and it is, of course, essential that the loftsman can provide
define a natural part of a surface, be the variables in the geometry facts for these discussions.
optimization problem. With this limited number of The criterion for a high-quality surface is FAN~A curves
variables, it is easier for the loftsman to define a well that are smooth. Thi.~ can be achieved by minimizing the
conditioned optimization problem, and it is also much strain energy of the FAN6A curves, as described in the
easier to interpret the optimization result. following section. An example of optimization is given
This is the way in which the loftsman works today, in Figure 10. The surface, part of a car roof, was originally
and the difference using optimization techniques is that defined as a physical model. Note again the corre-
the computer automatically finds the best solution for spondence between FANGA curves and reflection lines.
the given set of variables. Initially, only the chord-length variation FANGA curves
Another very important reason for defining surfaces (one curve per strip) were used in the optimization. The
in close interaction with an optimization function is that number of variables, controlling points and tangents was
the loftsman gets a clear picture of the geometry problem, ten. However some oscillations still remained on the
and is able to predict which combinations of constraints FANGA c u r v e s A 2 and A 3 after the optimiza'Jon. To
are geometrically possible. This is an important and very reduce these, an additional optimization was performed
difficult part of surface-definition work, since a surface modifying the P value and its derivative across section 3
practically never meets all the requirements of production, from the right. In this optimization, the FANGA curve
aerodynamics and other sources. A surface is always a A 2 w a s used.
[0P V[EN
CL
,qq
FI3
R2
R'[
J50 VIEN
Figure 7. P a r t o f f o r w a r d f u s e l a g e
OPTIMIZATION ALGORITHM AND This formulation leads to surfaces that are as smooth
IMPLEMENTATION and fair as possible for the given constraints. The
surface-shape type is not normally changed.
Two formulations of the objective function have been
In the second formulation, the objective is to minimize
implemented. Both have their advantages, and either may
the deviations between a set of target curves and
be preferred in different cases.
corresponding FANGA curves of the surface. The
In the first formulation, the curvature of the FANGA
deviations r are measured for N curves at M stations
curves is minimized. The curvature c of the FA~GA curves
along the surface strip(s).
created as cubic splines is computed for N curves at M
stations along the surface strip(s). N M
min F(.~)= ~" ~ (ru(~)) 2
N M
i=1 j=l
min F(~)= ~ ~ (cu(YO) z
i=1 j=l
subject to xl. . . . ~-X~'~Xupper" With target curves, the
subject to ~ . . . . ~<~ ~ X,pper, where the design variables x surface-shape type can be modified. To make the
are parameters controlling the geometric data, i.e. points, reflection lines in the car-roof example (Figure 10) change
tangents and P values, for defining the surface. from a convex to a concave shape would require target
A point can, for instance, be allowed to move along a curves.
line or a curve by letting one design variable control it. The minimization is performed as an iterative
Using two design variables, it can be moved in a plane. procedure in which the set of design variables x at each
A tangent can be allowed to move in a plane by letting step k are modified according to
one design variable control it. Using t w o design variables,
it can be moved in space. :~k ÷ 1 = YCk-- lkSR(yc)
s~(2) = G~ 1(2)gk(2)
Implementation
Figure 8. Part of fairing on control surface The optimization routines have been implemented on a
[Top figure: top view; bottom figure: ISO view.] VAX, and linked to our surface-definition system N M G .
Currently, the code is not portable, but our aim is to
convert our test program to a production (stand-alone)
v1 LI module that is robust, flexible and fast, and which can
be implemented in any geometry system.
CONCLUSIONS
Surface definition is today a very time-consuming process
which it would be very profitable to automate as much
as possible. However, it is also a task which is very
Figure 9. Input data to optimization function complex and can only be done by specialists.
[Point P~ defining the limit curve L 1 is allowed to move in the direction
Hence, our aim has been to find a method which
of VI; maximum deviations: - 0 . 1 to +0.5 mm. Tangent Tz defining
the limit curve L z is allowed to rotate towards V2; maximum deviations: can relieve the loftsman from time-consuming data
- 3 to +5' .] modification, but still work in close interaction with
him/her. This method we call FAN,A, and a great number
where 1k is a step-size parameter, and sk is the search of examples show that it is usable for both surface analysis
direction. and surface optimization.
Our optimization approach leads to very well
conditioned and small minimization problems. All our
Search direction tests also indicate that the time taken for the definition
The evaluation of search directions is dependent on the of high-quality surfaces can be drastically reduced with
optimization method used. the FANGA optimization method.
STRIP I
To v
STRIP 2
~IRTP
NON-OPTIMIZED OPTIMIZED
STRIP I STRIP 1
CL CL
Al Rl
R2 R2
R3 A3
R4 A~
STRIP 2 STRIP 2
- CL CL
A1 A1
R2 A2
---" A3 R3
RLI ALl
STRIP 3
CL STRIP 3
CL
RI
,--..._..
Rl
A2
"•A2 -------- A3
ALl
R3
(a) ~R L~
NON-OPTIMIZED OPTIMIZEO
j,
-I t I Ill I
/ I / / / . . . . I ~ ~
I
\ \ I / \ \--I/ -~": \ I / ;
I I II Iit lib--...... I I / r
I I I I IIII II\"." ~ ' - ~ , I t t I
"7TT
(c) I! lIlltfltt l
F i g u r e 10. Car roof," (a) FANGA optimization, ( b ) reflection lines, ( c ) principal curvature directions