Mining Sequence Patterns from Wind Tunnel Experimental Data

1
Mining Sequence Patterns from Wind Tunnel
Experimental Data

• Zhenyu Liu, Wesley W. Chu, Adam Huang, Chris
  Folk, Chih-Ming Ho
• wwc,vicliu_at_cs.ucla.edu, pohao,chrisf,chihming_at_
  ucla.edu
• Computer Science Department Mechanical and
  Aerospace Engineering Department
• University of California
• Los Angeles, California

2
Outline

• Problem statement
• Scientific experimental data characteristics
• Conventional mining methods
• Decision tree
• Association rules
• Mining of sequence patterns
• Conclusion

3
Delta Wing Aircraft Control via MEMS Actuators

• Vortices symmetry is broken by the actuation of
  MEMS actuators, resulting in desirable
  aerodynamics loadings

Werle, 1958
4
Problem Statement

• Inputs angle of attack, stream velocity,
  actuation angle
• Output rolling moment
• Problem discover knowledge on input-output
  relationship

5
Scientific Experimental Data Characteristics

• The output is highly dependent on all inputs. A
  subset of inputs is inadequate to predict the
  output.
• Sequences of input-output relationships contained
  (e.g. the rolling moment w.r.t. the actuation
  angle)

6
Conventional MethodsDecision Tree Generation

• High coverage but low accuracy (an error rate of
  46.35 in predicting the original dataset using
  the decision tree)
• Reason
• The decision tree generation algorithm uses
  univariate-split strategy to induce the
  input-output relationship
• Each single input has low prediction power over
  the output.

Angle ofAttack
0,5,10
15,20,25,30,35
MZ in-0.00024,0.00028
ActuationAngle
60,100
40,80,120,140
MZ in-0.0124,0.00537
Angle ofAttack
15,20
25,30,35
MZ in-0.00024,0.00028
MZ in-0.01179,-0.0038
7
Conventional Methods Rule Induction

• Acceptable accuracy
• Low input state space coverage (which is 25) and
  insufficient for flight control applications

8
Conventional Methods The Cause of The Low
Coverage

• The output variable in scientific dataset cannot
  be summarized using a subset of the inputs
• Rules induced from the sensitive input regions
  (large angle of attack in this case) cannot have
  both high confidence and high support

9
Mining of Sequence Patterns

• Extract sequences from the dataset
• Bottom-up sequence clustering (binary hierarchy)
  based on Euclidean distance measure
• Sequence pattern extraction from the binary
  cluster hierarchy based on the variance measure
• Rule induction from sequence patterns

10
Mining of Sequence Patterns1. Sequence
Extraction

• Merging the output with one of the input to form
  a composite output variable. More specifically
• A dataset D with inputs X1, , Xn and an output Y
• A predicate p defined on the inputs
• A sequence of Y w.r.t. Xi (1? i? n) characterized
  by p is a set of 2-item tuples lty1, xi1gt, ,
  ltym, ximgt calculated by ? Y, Xi(?p(D))

a sequence characterized byp aoa20, vel10
11
Mining of Sequence Patterns 2. Bottom-up
Sequence Clustering

• Using the Euclidean distance measure to generate
  a binary cluster hierarchy

12
Mining of Sequence Patterns 3. Sequence Pattern
Extraction

• From the hierarchy, merge branches with variances
  below a user-specified threshold (0.35 in this
  example)

13
Mining of Sequence Patterns 4. Rule Induction
from Sequence Patterns

• Cluster 8 as an example
• Sample rules generated
• IF angle of attack 35? THEN the rolling moment
  curve with actuation angle follows mean(cluster
  8), confidence 100, variance 0.243029
• Rules have higher coverage and confidence, and
  the accuracy (of the mean) is controllable
  through the variance measure

cluster 8

wvar
0.243029

aoa 35 vel 10

aoa 35 vel 15

aoa 35 vel 20

The average of sequences in cluster 8, or,
mean(cluster 8)
All sequences in cluster 8
14
Conclusion

• Developed a mining technique to discover
  relationship for highly correlated input-output
  pairs
• Conventional methods (decision tree or rule
  induction) fail to generate knowledge with both
  high coverage and accuracy
• Developed a new technique for mining sequence
  patterns from the wind tunnel experimental data
• Sequence extraction
• Sequence clustering (binary hierarchy) based on
  the Euclidean distance measure
• Sequence patterns extracted from the binary
  cluster hierarchy using variance measure
• Rule induction from sequence patterns
• Rules generated
• Nontrivial to experimenters
• Useful for flight control

15
Directly Applying Conventional Mining Methods

• Conventional methods works with categorical
  variables.
• The first step will be a discretization of the
  output variable
• 6 partitions generated -0.01179, -0.00380,
  -0.00380, -0.00138, -0.00138, -0.00024,
  -0.00024, 0.00028, 0.00028, 0.00124,
  0.00124, 0.00537
• Then apply decision tree or rule induction to the
  discretized dataset