The document discusses gradient-based fault isolation techniques. It begins by explaining the need for both fault detection and fault isolation in industrial processes. It then discusses commonly used residual-based approaches for fault detection and introduces the motivation for gradient-based fault isolation approaches. These approaches use partial derivatives of models to identify the process variables most responsible for detected faults. The document outlines some tools for visualizing fault isolation results and discusses opportunities to improve gradient aggregation methods and further narrow isolation performance gaps. It concludes that gradient-based techniques can perform fault isolation without prior fault information and more development is still needed to strengthen results.
1. Gradient-based Fault Isolation
Residual-based Fault Detection Systems
Francisco Serdio Fernández
Department of Knowledge-Based Mathematical Systems
Johannes Kepler University Linz, Austria
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
for
http://www.flll.Francisco Serdio jku.at/staff/francisco
2. Why Fault Detection (FD) ?
Why Fault Isolation (FI) ?
FD with Residual-based approaches
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more ?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
3. WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Why Fault Detection?
http://www.flll.Francisco Serdio jku.at/staff/francisco
4. Why Fault Detection?
Products with high quality demands
High quality is required also in the production chain
High quality is required also in the supply chain
[1] D. Blanchard. Supply Chain Management Best Practices. John Wiley & Sons,
Hoboken, NJ, USA, 2007.
Continuity in the production lines
Minimum down-time
[2] R. Iserman. Fault-Diagnosis Applications. Model-Based Condition Monitoring:
Actuators, Drives, Machinery, Plants, Sensors, and Fault-tolerant Systems. Springer,
Berlin Heidelberg, Germany, 2011.
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
High quality processes imply
http://www.flll.Francisco Serdio jku.at/staff/francisco
5. WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Why Fault Detection?
http://www.flll.Francisco Serdio jku.at/staff/francisco
6. Manual supervision is not affordable or in some
cases simply impossible
The precision of manual supervision usually depends
on the experience of the operators
and even on their performance on a given day
[3] E. Lughofer, J.E. Smith, P. Caleb-Solly, M. Tahir, C. Eitzinger, D. Sannen and M.
Nuttin. (2009). Human-machine interaction issues in quality control based on on-line
image classication. IEEE Transactions on Systems, Man and Cybernetics, part A:
Systems and Humans, 39(5), 960-971.
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Why Fault Detection?
Manual process supervision
http://www.flll.Francisco Serdio jku.at/staff/francisco
7. Why Fault Detection (FD) ?
Why Fault Isolation (FI) ?
FD with Residual-based approaches
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more ?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
9. Why Fault Isolation?
Multiple sensor networks turned out to emerge
in industrial settings and factories
Huge amount of sensors and actuators to check
Manual supervision is not affordable or in some
cases simply impossible
Any valuable information regarding where the fault
is
located could be a great aid for the operator
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Isolation !
http://www.flll.Francisco Serdio jku.at/staff/francisco
10. Why Fault Detection (FD) ?
Why Fault Isolation (FI) ?
FD with Residual-based approaches
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more ?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
11. FD with Residual-based approaches
Algebraic relationships among different sensors
Difference relationships among different sensor
outputs and actuator inputs
Inconsistencies, expressed as residuals, can be
used for detection and isolation purposes
[4] V. Venkatasubramanian, R. Rengaswamy, S. Kavuri and K. Yin. (2003). A review of
process fault detection and diagnosis: Part iii: Process history based methods.
Computers & Chemical Engineering, 27(3), 327-346.
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Analytical Redundancy
Direct redundancy
Temporal redundancy
http://www.flll.Francisco Serdio jku.at/staff/francisco
12. FD with Residual-based approaches
Analytical Redundancy graphically
Moving from the signal space to the residual space: illustrating an untypical signal pattern
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
13. Tracking residuals within a dynamic tolerance band
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
14. Recall FD with Residual-based
approaches
More information regarding Fault Detection in
[5] F. Serdio, E. Lughofer, K. Pichler, T. Buchegger and H. Efendic, Data-Driven Residual-Based
Fault Detection for Condition Monitoring in Rolling Mills. Proceedings of the IFAC Conference on
Manufacturing Modeling, Management and Control, MIM '2013, St. Petersburg, Russia, 2013, pp.
1546-1551. (Winner of MIM 2013 Best paper award)
[6] F. Serdio, E. Lughofer, K. Pichler, T. Buchegger, and H. Efendic, Residual-based Fault Detection
using Soft Computing techniques for Condition Monitoring at Rolling Mills. Information Sciences,
vol. 259, pp. 304–330, 2014.
[7] F. Serdio, E. Lughofer, K. Pichler, T. Buchegger, M. Pichler and H. Efendic, Multivariate Fault
Detection using Vector Autoregressive Moving Average and Orthogonal Transformation in the
residual Space. Annual Conference of the Prognostics and Health Management Society, PHM 2013,
New Orleans, LA, USA, 2013, pp. 548-555.
[8] F. Serdio, E. Lughofer, K. Pichler, T. Buchegger, and H. Efendic, Fault Detection in Multisensor
Networks based on Multivariate Time-series Models and Orthogonal Transformations.
Information Fusion, vol. under revision (minor), 2014.
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
15. Why Fault Detection (FD) ?
Recall FD with Residual-based approaches
Why Fault Isolation (FI) ?
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
16. Motivation of the FI Gradient-based
approaches
We are blind about faults
We do not know how a fault looks like
We do not have fault patterns (labeled data)
Process variable contribution plot
There is an extension to non-linear PCA
It reverts back to the original process variables
[9] P. Miller, R. Swanson, and C. Heckler, Contribution plots: A missing link in multivariate quality
control. Applied Mathematics and Computer Science, vol. 8, p. 775792, 1998.
[10] F. Jia, E. Martin, and A. Morris, Nonlinear principal components analysis with application to
process fault detection. International Journal of Systems Science, vol. 31, p. 14731487, 2001.
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
There is literature about PCA
http://www.flll.Francisco Serdio jku.at/staff/francisco
17. Motivation of the FI Gradient-based
approaches
Partial derivatives !
With respect to a specific dimension can indicate the
relative importance of the corresponding variable
(channel) on that function
Can be computed according to the model expression
Can be computed by means of numeric tricks
We can plug a FI system to any FD model !
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
18. Motivation of the FI Gradient-based
approaches
How do we revert back to the original process
variables?
We compute the gradients of the model variables
We aggregate the gradients
We get a candidate responsible variable
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
We take the warning models
Crisp decision
Fuzzy decision
http://www.flll.Francisco Serdio jku.at/staff/francisco
19. Aggregating gradients
Biggest gradient as faulty channel
A channel is either (properly) isolated or not
Several channels are proposed as faulty
There are normalized against the channel with the
highest gradient aggregation
By definition, it will produce always better results than
its crisp counterpart
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Crisp decision
Winner takes all approach
Fuzzy decision
http://www.flll.Francisco Serdio jku.at/staff/francisco
20. Why Fault Detection (FD) ?
FD with Residual-based approaches
Why Fault Isolation (FI) ?
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more ?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
21. Tools to depict Fault Isolation
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
22. Why Fault Detection (FD) ?
FD with Residual-based approaches
Why Fault Isolation (FI) ?
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
23. WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Results
http://www.flll.Francisco Serdio jku.at/staff/francisco
24. WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Results
http://www.flll.Francisco Serdio jku.at/staff/francisco
25. Why Fault Detection (FD) ?
FD with Residual-based approaches
Why Fault Isolation (FI) ?
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more ?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
26. Can do we more ?
We must work in how to aggregate the
gradients
Weight the gradients with other data
We are using violation degree of the threshold
We are using quality of the model
Goal: narrow the Fault Isolation Gap (FIG)
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Time frames (sliding windows)
http://www.flll.Francisco Serdio jku.at/staff/francisco
27. Why Fault Detection (FD) ?
FD with Residual-based approaches
Why Fault Isolation (FI) ?
Motivation of the FI Gradient-based approaches
Tools to depict Fault Isolation
Results
Can do we more ?
Conclusions
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
http://www.flll.Francisco Serdio jku.at/staff/francisco
28. Conclusions
We can perform Fault Isolation (FI) without
information about the faults
Only based on warning models and gradients
We have introduced new tools to depict FI
We must still strength the results
WCCI 2014 / July 6-11 / Beijing, China
francisco.serdio@jku.at
Graphically
Numerically
http://www.flll.Francisco Serdio jku.at/staff/francisco
29. Thanks a lot for your attention!
WCCI 2014 / July 6-11 / Beijing, China
{francisco.serdio,edwin.lughofer}@jku.at
http://www.flll.jku.at/staff/{Francisco Serdio, Dr. Edwin Lughofer francisco,lughofer}