Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

INTRODUCTION TO

EXPERIMENTAL DESIGN

Lars Drugge

KTH Vehicle Dynamics

Testing strategies

Problems occur which are dependent on several different factors,


some of these are:
• controllable
• observable
• unknown

Natural testing strategy, “one-factor-at-a-time”:


• Each variable is studied individually, while all others are held constant
• Very time consuming
• Information about possible interactions is lost

2
Design of Experiments

Multivariable testing methods – factorial experimental design:


• Several variables are varied simultaneously
• Shorter testing time
• Interactions between variables can be quantified
• Requires more planning, but gains greater efficiency

1. 2.

4. 3.

Example: Airbag test

• Development of airbags:
- Investigate the influence of airbag design on the
chest acceleration of a crash test dummy

• Three factor test with two levels for each factor:

Inflation speed S Low (-) High (+)


Airbag volume B Small (-) Large (+)
Vent area V Small (-) Large (+)

4
Test plan and results

Balanced test plan:


• Each factor is tested the same number of times (4) on
each level (2), and no test is repeated
Vent area constant

Trial

Subset

Subset of results
Trial

Denominator

Denominator = number of times each factor has been tested at respective level

Effect, ci = difference between the average value of all the results for i (+)
and the corresponding average value of the results for i (-)

6
Effect of factor S
Trial

Denominator

55.6  52.2 56.2  61.6


Effect of factor S: cs  
2 2
 56.2  55.6  61.6  52.2
Rewriting: cs   5.0m / s 2  S
2

Effect of factor B
Trial
Result vector

Coefficient vector
Denominator

Numerator = sum of the element-wise product of the coefficient vector


and the result vector

 56.2  55.6  61.6  52.2


Effect of factor B: cB   1.0m / s 2  B
2

8
Graphical representation

Inflation speed, S Airbag volume, B

Interaction effect
Trial

Denominator

Interaction effect, cS*B = half the difference between the effect of S when
B = B(+) and the effect of S when B = B(-)

(52.2  61.6)  (55.6  56.2)


cS  B   4.4m / s 2  SB
2
10
Interaction effect
56.2  55.6  61.6  52.2
Rewriting: cS  B   4.4m / s 2  SB
2

Coefficent vector = element-wise product of the coefficient vectors for


S and B

11

Evaluation of subset of results


Orthogonal L4 design matrix
Trial

Denominator

• Inflation speed, S, is more important than airbag volume, B


• The interaction between them control the choice of airbag volume

12
Orthogonal design matrix

• Characterised by the fact that the scalar product of


each pair of column vectors is zero

• Lists of orthogonal test matrices available

• New test matrices can be generated:


for factors with two, three or more levels, as well as
combinations where some factors are tested on two
levels and others on three levels

13

Result of orthogonal L8 matrix


Interaction effects
Introduce factor V

Trial

• =

Denominator

14
Pareto chart

• Absolute values of effects are arranged in order of magnitude


• Random experimental variations should appear as a straight line

15

Prediction

The effects can also be used to predict the outcome of coming tests:

The prediction = ỹ + ½ . the scalar product of the Coefficient row and the Effect row.

Prediction => 55.1 + ½ (5.8 – 2.35 + 2.55 – 3.5) = 56.3 m/s2

Trial 1 = 56.2 m/s2


16
Full factorial design

A full factorial design where k factors are tested on p levels


requires that the total number of experiments N are

N = pk

k = 3, p = 2 gives N = 23 = 8

k = 7, p = 2 gives N = 27 = 128

• The number of trials grows exponentially with the number of


factors!

17

Example: Vehicle handling

• Evaluate the influence of vehicle load and tyre pressure on


vehicle handling

• Analyse the vehicle steady state handling characteristics

• Determine the understeer gradient, Kus

18
Understeer gradient

• If Kus = 0 the vehicle is neutralsteered. This implies


that the steering wheel angle is constant, independent
on speed, when driving in a circle with radius R.

• If Kus > 0 the vehicle is understeered. As the speed increases


an increasing steering wheel angle is needed to keep the
vehicle on the circular path.

• If Kus < 0 the vehicle is oversteered. A decreasing steering


wheel angle is needed to keep the vehicle on the circular
path as the speed increases.

19

Experimental design

• Two factors (k=2) are tested on two levels (p=2):

Factor Low level High level

Axle load (rear), Fz [kg] 0 183


Tyre pressure (rear), P [bar] 1.8 2.2

• Number of trials:
N = pk = 22 = 4
=> L4 orthogonal full factorial test matrix

20
Test plan and results
Trial Fz P Fz x P Kus
1 -1 -1 1 5.87 .10-3
2 1 -1 -1 4.35 .10-3
3 -1 1 -1 5.85 .10-3
4 1 1 1 4.83 .10-3

-1.28 .10-3 0.24 .10-3 0.26 .10-3 5.22 .10-3

Pareto chart
0.0014
0.0012
0.001
0.0008
0.0006
0.0004
0.0002
0
Fz P xxFz
Fz P P

21

Graphical representation
Kus

High pressure

Low pressure

Fz
High Axle load Low

22
Graphical representation
Kus

Low axle load

High axle load

P
High Pressure Low

23

Fractional factorial design

The principle of fractional factorial design:

• Give up the possibility to analyse one or more


interactions

• The corresponding column can instead be used for


an additional main factor

24
L8 matrix with four factors

Additional factor, D

c1. c2 = c3 c3 = c4. c7
25

Full factorial design

Full factorial design with 7 factors on two levels

N = 27 = 128

26
Fractional factorial design

1/2 Fractional Factorial Experiment


1/4
1/8

N = 16
64
32

(Trial boxes
boxes marked
marked in
inblack)
black)
27

Resolution
Resolution = Ability to isolate main effects from two-
and multiple-factor interactions

• Resolution V:
- All main factors and two-factor interactions are isolated
from other two-factor interactions

• Resolution IV:
- All main factors are isolated from two-factor interactions,
while two-factor interactions can be mixed with other
two-factor interactions

• Resolution III:
- Main factors and two-factor interactions can be mixed

28
Example of resolutions

Resolution
Matrix III IV V
L8 7 4 3*
L16 15 8 5
L32 31 16 6
L64 63 32 8

* Full factorial design

29

Resolution for an L16 matrix


Additional factor, E

Additional factors, E & F

30
Example: Roll-over stability

Increase knowledge regarding heavy vehicle


roll-over stability:

• For increased passive safety

• When developing active roll-over prevention


systems

Source: E. Dahlberg; ”Commercial Vehicle Stability - Focusing on Rollover”,


Doctoral Thesis, KTH Vehicle Dynamics, 2001.

31

Roll-over threshholds

Steady state roll-over threshhold (SSRT)


• Maximum lateral acceleration

Dynamic roll-over threshhold (DRT)


• Worst case measure (kinetic energy)
=> minimum lateral acceleration

32
Vehicle parameters
• Five factors on three levels (low, intermediate, high)
• Parameters set by the vehicle manufacturer
• Simulation using MBS model in ADAMS

33

Orthogonal test matrix

N = 35 = 243

Fractional factorial
design, L27
=> 27 tests

34
Results – SSRT vs DRT
Linearised effects

35

Non-linear effects - DRT


Non-linear main effects Non-linear interaction effects

36

You might also like