Automatic Fault Detection in Friction Stir Welding

1
Automatic Fault Detection in Friction Stir Welding

• Paul Fleming
• Thomas Bloodworth
• Tracie Prater
• David Lammlein
• George E. Cook
• Alvin Strauss
• D. M. Wilkes
• David Delapp
• Thomas J. Lienert
• Matt Bement

2
Friction Stir Welding

• Recently (1991) developed solid state welding
  technique
• Uses mechanical stirring to join metals
• Yields high weld strength
• Can be used to join aluminum

3
Defects in FSW

• Defects can occur in FSW and undermine the
  integrity of the weld
• Detecting the presence of these defects is
  important for ensuring quality welds
• Defects include
• Worm Hole Faults
• Gaps
• Tracking errors

4
Goal

• Develop means of non-destructively detecting
  fault occurrence, recognizing fault type and
  correcting in real time.
• Predictive Process Manufacturing

5
Goal
6
Experiments

• Conducted initial research into two fault types
• Gap faults in lap welds
• Tracking errors in butt welds
• Will present gap faults today

7
Gap Fault Experiment

• Friction Stir Lap Welding is FSW performed by
  placing samples one on top of the other and
  plunging the rotating tool through the first
  sample in to the second to form the joint
• Gaps can exist between the samples when the
  fit-up is not perfect and possibly lessen the
  weld strength

8
Gap Fault Experiment

• Gaps were milled into samples at depths ranging
  from
• .0002 - .005

9
Materials and Equipment

• 6061 Aluminum Sample
• 01 Steel Tool
• Spindle at 2000 RPM
• Traverse at 16 IPM

10
Sensing

• Kistler Dynamometer

11
Results

• Results indicate that this method can be used to
  detect even small (ten-thousandths of an inch)
  gaps in FSLW
• Machine Learning techniques for finer detection

12
Results Axial Force
13
Results

• A noticeable drop in axial force is shown when
  gaps are over .002
• For smaller gaps, can attempt to use information
  in the frequency of the axial force combined with
  PCA, LDA and SVM to discover gap presence

14
PCA

• Principal Component Analysis
• Defines new set of axis along directions of
  maximum variance
• Often used as a means of dimensionality reduction
• When applied to lap weld data...

15
PCA
16
LDA

• Linear Discriminant Analysis
• Similar to PCA but axis are selected to maximize
  between-class scatter and minimize within-class
  scatter

17
LDA
18
SVM

• Support Vector Machine
• Attempts to find optimal dividing plane between
  classes in high dimension
• Used in this work as a prediction algorithm where
  it is trained on one subset of data and asked to
  predict the classification of the remaining data

19
Accuracy of SVM
20
SVM Regression

• Modification where output is function estimate
  rather than classification

21
SVM Regression
22
Conclusion

• Using FSLW as a testbed, we demonstrate the use
  of Machine Learning in developing fault detection
  systems which can non-destructively detect fault
  occurrence as part of a PPM operation.

23
References

• George E. Cook, Reginald Crawford, Denis E.
  Clark, and Alvin M. Strauss. Robotic friction
  stir welding. Industrial Robot, 31(1)5563,
  November 2004.
• Terry Khaled. An outsider looks at friction stir
  welding. Technical report, Federal Aviation
  Administration, 2005.
• Ericsson, M., Jin, L.-Z. Sandstrom, R. (2007),
  Fatigue properties of friction stir overlap
  welds, International Journal of Fatigue 29,
  5768.
• Fukunaga, K. (1972), Introduction to Statistical
  Pattern Recognition, Academic, New York.
• Mishina, O. K. Norlin, A. (2003), Lap joints
  produced by fsw on ?at alu minum en aw-6082
  pro?les, in 4th International Symposium on
  Friction Stir Welding.
• Shlens, J. (2005), A tutorial on principal
  component analysis, Technical report, Systems
  Neurobiology Laboratory, University of
  California, San Diego, La Jolla, CA.

24
Acknowledgments

• This work supported by
• AWS
• LANL
• NASA