Title: Pattern Recognition Approach for Fault Diagnosis of DAMADICS Benchmark
1Pattern Recognition Approach for Fault Diagnosis
of DAMADICS Benchmark
- Cosmin Bocaniala
- University Dunarea de Jos from Galati, Romania
- Andrzej Marciniak
- University of Zielona Gora, Poland
- Jose Sa da Costa
- Instituto Superior Tecnico, Lisbon, Portugal
2...
- Andrzej Marciniak contribution
3A Novel Fuzzy Classifier
- fault diagnosis may be seen as a classification
problem - building a map between symptoms space and the set
of faulty states - two main advantages of the developed fuzzy
classifier - the high accuracy with which it distinguishes
between different categories - the fine precision of discrimination inside
overlapping zones
4Previous work
- three main directions of using fuzzy classifiers
- neuro-fuzzy systems
- robust to uncertainties and noise
- collections of fuzzy rules
- transparent symptoms-faults relationships via
linguistic terms - represent normal state and each faulty state as
fuzzy subsets of the symptoms space
5Point-to-point similarity
- the similarity s(u,v) between two points is
computed using a dissimilarity measure d(u,v) - the ß parameter plays the role of a threshold
value for the similarity measure - single or hybrid similarity measures
6Point-to-set similarity
- the similarity measure between two points can be
extended to a similarity measure between a data
point and a set
7Induced fuzzy sets
- the fuzzy membership functions are induced by the
point-to-set affinity - each category has associated a different ß
parameter
8Computational aspects
- the main computational issue is the search for
the set of parameters that provide the best
performance - genetic algorithms (slow)
- hill climbing (fast)
- particle swarm optimization (best!)
9GA vs HC
No. exp Initial Final No.calls (Method) fitness fi
tness classifier
1 (GA) 138.83 147.98 340 2 (GA) 140.40 149.54 340
3 (GA) 144.97 151.26 340 4 (GA) 140.83 149.82 340
5 (GA) 139.98 151.08 340 1 (HC) 138.83 149.54 136
2 (HC) 140.97 150.59 141 3 (HC) 142.98 151.50 108
4 (HC) 141.90 151.30 124 5 (HC) 137.90 149.78 106
10GA vs PSO
No. exp Initial Final No.calls (Method) fitness fi
tness classifier
1 (GA) 138.83 147.98 340 2 (GA) 140.40 149.54 340
3 (GA) 144.97 151.26 340 4 (GA) 140.83 149.82 340
5 (GA) 139.98 151.08 340 1 (PSO) 124.18 151.25 100
2 (PSO) 121.62 150.03 220 3 (PSO) 117.65 146.46 1
60 4 (PSO) 127.47 151.38 140 5 (PSO) 134.07 156.30
140
11Results on DAMADICS benchmark Step I
- the effects of six out of the 19 faults on this
set of sensor measurements are not
distinguishable from the normal behavior, F4,
F5, F8, F9, F12, F14 - also, there can be distinguished three groups of
faults, F3, F6, F7, F10, and F11, F15, F16,
that share similar effects on the measurements
and, therefore, can be easily confound with
faults in the same group.
12Results on DAMADICS benchmark Step I
The large overlapping between F3 and F6
13Results on DAMADICS benchmark Step I
The large overlapping between F7 and F10
14Results on DAMADICS benchmark Step I
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
F1 N N N N N N N N N F1 F1 F1 F1 F1 F1 F1 F1 F1 F1 F7
F2 N N N N F2 F2 F2 F2 F2 F2 F2 F19 F2 F2 F2 F2 F2 F2 F2 F2
F3 N N N F6 N N F6 F3 F3 F3 F1 F3 F6 F3 F3 F3 F3 F3 F3 F3
F6 N N N N N N N N N F6 N F6 F6 F6 F6 F6 F6 F6 F6 F3
F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F7 F1
F10 N N N N N F15 F10 F1 F1 F10 F10 F10 F10 F10 F1 F7 F7 F7 F7 F7
F11 N N N N N N F11 N F11 N F11 F11 F11 F11 F11 F11 F11 F11 F11 F11
F13 N F3 F3 F3 F3 F3 F13 F13 F13 F13 F13 F13 F13 F13 F13 F13 F13 F13 F13 F13
F15 N F15 N F15 N N N F15 F15 F15 N F15 F15 F15 F15 N N F15 N F15
F16 N N N N N N N N N N N N N N F16 N F16 F15 F16 F16
F17 N N N N F17 F17 F17 F17 F17 F17 F17 F17 F17 F17 F17 F17 F17 F17 F17 F17
F18 F3 F3 F3 F3 N F18 F18 F18 F18 F18 F18 F18 F18 F18 F18 F18 F18 F18 F18 F18
F19 N N F2 F2 F2 F2 F2 F19 F19 F19 F19 F19 F19 F19 F19 F19 F19 F19 F19 F19
15Conclusions
- advantages high accuracy discrimination between
different categories, and fine precision inside
overlapping zones - fast parameters tuning using PSO
- good performances on the DAMADICS benchmark, Step
I