Sei sulla pagina 1di 5

Bivariate Quadratic Regression and Conics

April 27, 2011

John Alexiou - Pr. Analytical Engineer


Koyo Bearings USA, Greenville, SC
May 2010

1 Introduction
In designing a valvetrain rocker arm, with predefined valve motion caution is needed that the resulting
camshaft shape follows the constraint that its minimum radius of curvatue be more than a fixed amount.
Two main design variables affect the minimum cam curvature. One was the cam base circle and the other
was the cam side rocker arm length. In Figure 1 the two quantities of variation “Cam Side Arm” and “Cam
Base Radius” are shown. The goal of this paper is to show the mathematical relationship between the two
inputs (the design variables) and the one output (the radius of curvature) and how to best choose these
variables to meet our goal.
The resulting curvature is formulated mathematically as

 x1 is cam base radius
crv = f (x1 , x2 ) where x2 is cam side arm length (1)
crv is the resulting curvature

Figure 1: Sketch of two quantities (x1 , x2 ) to vary in order to optimize the minimum cam curvature (crv).

1
Dynamics John Alexiou - Pr.Analytical Engineer

2 Measurements & Model


The baseline values for (x1 , x2 ) is (10.5 mm, 21.422 mm) resulting in a minimum negative curvature of
crv = 58.8 mm. In all Table 1 shows the data generated from multiple runs in a mini DOE. All this data
can be fitted to a bivariate quadratic regression model and solved for 6 coefficients

crv = C1 + C2 x1 + C3 x21 + C4 x2 + C5 x22 + C6 x1 · x2 (2)

To do the regression the problem is formulated in a matrix system of equations

Y = A·C (3)
 
C1
.. ..
   
 C2 
. .  
    C3 
 crvi 
 1
=  x1 x21 x2 x22 x1 x2  
  (4)

..

..

 C4 

. .  C5 
C6

Y A
z
 }| { z
 }| {
58.8 1.0 10.5 110.0 21.4 459.0 225.0

 81.1  
  1.0 12.5 156.0 21.4 459.0 268.0 


 124.0  
  1.0 14.5 210.0 21.4 459.0 311.0 

 73.2   1.0 10.5 110.0 20.4 417.0 214.0 
    C

 82.1  
  1.0 11.2 125.0 20.4 417.0 229.0 z
 }| {

 68.9  
  1.0 10.5 110.0 20.7 427.0 217.0 
 C1

 76.5  
  1.0 10.5 110.0 20.3 410.0 213.0 
 C2 

 78.3  
= 1.0 10.5 110.0 20.1 404.0 211.0  C3 
   (5)

 79.5  
  1.0 10.5 110.0 19.9 397.0 209.0 
 C4 


 61.8  
  1.0 10.5 110.0 21.2 448.0 222.0 
 C5 

 65.1  
  1.0 10.5 110.0 20.9 438.0 220.0 
 C6

 74.5  
  1.0 11.0 121.0 20.7 427.0 227.0 


 88.7  
  1.0 12.0 144.0 20.7 427.0 248.0 


 78.5  
  1.0 12.0 144.0 21.2 448.0 254.0 

 91.4   1.0 12.5 156.0 20.9 438.0 262.0 
105.0 1.0 13.5 182.0 21.2 448.0 286.0
which has a solution of

C = (AT A)−1 (AT Y ) (6)


 
−182.69
 28.37 
 
 2.486 
C = 
 20.19 
 (7)
 
 0.0350 
−3.466

The error of this fit can be gaged by the mean


s
Pn 2
(Y − AC)
i=1
e = (8)
n
= 0.533 (aprox. 0.65%) (9)

2
Dynamics John Alexiou - Pr.Analytical Engineer

3 Prediction & Solution Families


So the predicted curvature crv given (x1 , x2 ) is
crv = 2.49 x1 2 − 3.47 x1 x2 + 28.4 x1 + 0.035 x2 2 + 20.2 x2 − 182.7 (10)
This forms a conic section in (x1 , x2 ) described by the conic matrix equation
C3 C26 C2
  
2 x1
x1 x2 1  C26 C5 C4

  x2  = 0 (11)
2
C2 C4 1
2 2 C1 − crv
  
 2.486 −1.733 14.185 x1
x1 x2 1  −1.733 0.035 10.095   x2  = 0 (12)
14.185 10.095 −182.69 − crv 1

Iteration x1 x2 crv
0 10.50 21.422 58.8
1 12.50 21.422 81.1
2 14.50 21.422 123.9
3 10.50 20.422 73.2
4 11.20 20.422 82.1
5 10.50 20.672 68.9
6 10.50 20.255 76.5
7 10.50 20.088 78.3
8 10.50 19.922 79.5
9 10.50 21.172 61.8
10 10.50 20.922 65.1
11 11.00 20.672 74.5
12 12.00 20.672 88.7
13 12.00 21.172 78.5
14 12.50 20.922 91.4
15 13.50 21.172 105.2

Table 1: Curvature measurements for various base circles and cam side arms lengths.

4 Optimization
To move the design variables (x1 , x2 ) such that a particular curvature crv = 100 is achieved first we need to
find the set of points that satisfy this condition. Then decide which path will move the design point onto the
desired curve with the least amount of change in the parameters. Figure 2 shows this process graphically.
The original point is located at (x1 , x2 )=(10.5, 21.422) and it needs to move to reach the curve defined
by
100 = 2.49 x1 2 − 3.47 x1 x2 + 28.4 x1 + 0.035 x2 2 + 20.2 x2 − 182.7 (13)
The distance moved is defined by
p
d= (x1 − 10.5)2 + (x2 − 21.422)2 (14)
To solve the optimization problem one needs to minimize d while still maintaining the equation of the curve.
Typically the curve equation is solved for one variable x1 and substituted into d. Then the derivative with
respect to x2 is taken and set to 0 and solved for x2 . The solution for x2 is used to solve for x1 . There is no
analytical solution at this point and some numerics need to be deployed.

3
Dynamics John Alexiou - Pr.Analytical Engineer

Figure 2: Optimization routine moved the design point from the original curve (Gray) to the desired curve
(Red) with the least amount of change.

5 Conic Transformation
The problem may be simplified a little bit by using a coordinate transformation to decompose the design
points along the primary axes of the conic section (focus and directrix). The general transformation from
(x1 , x2 ) to (u, v) is
     
x1 cos ψ − sin ψ c1 u  c1 is the center along the x1 axis
 x2  =  sin ψ cos ψ c2   v  where c2 is the center along the x2 axis (15)
1 0 0 1 1 ψ is the orientation angle

This transformation diagonalizes the conic matrix when the following constants are used
2 C2 C5 − C4 C6
c1 = = 6.170 (16)
C62 − 4 C3 C5
2 C3 C4 − C2 C6
c2 = = 17.04 (17)
C62 − 4 C3 C5
 
π 1 C3 − C5
ψ = − arctan = 1.09316 (18)
4 2 C6

yielding the simplified conic equation


  
 K1 u
u v 1  K2  v  = 0 (19)
K3 1

4
Dynamics John Alexiou - Pr.Analytical Engineer

with
 
1
q
K1 = C3 + C5 + C62 + (C3 − C5 )2 = 0.86193 (20)
2
 
1
q
K2 = C3 + C5 − C62 + (C3 − C5 )2 = −3.3823 (21)
2
C 2 C5 − C2 C4 C6 + C3 C42
K3 = C1 − crv + 2 = 23.1333 (22)
C62 − 4 C3 C5

6 Optimization Solution

Figure 3: Transformed optimization of design such that minimum curvature is 100 mm.

The transformed solution curve is now


 u 2  v 2
− + =1 (23)
5.1807 2.6150
and the goal is to minimize the distance
p
d= (u − uc )2 + (v − vc )2 (24)
from the original design point (uc , vc ) define as
   −1    
uc cos ψ − sin ψ c1 10.5 5.8826
 vc  =  sin ψ cos ψ c2   21.422  =  −1.8302  (25)
1 0 0 1 1 1
The optimization solution is    
ud 4.223
 vd  =  −3.374  (26)
1 1
which corresponds to the design point (x1 , x2 )=(11.108, 19.2386)

Potrebbero piacerti anche