Data analysis of aircraft
engines
Tomáš Rudolecký
Outline
• Why we want to use Machine Learning techniques
for engine fault prediction
• Characteristics of the data
• Possible approaches
• Examples
Engine critical systems
Short-term change
is detected.
Long-term change
is detected.
• The objective in change detection is to produce evidence for HI mapping. Short-term
CD implements a WE-type rule to detect statistically significant shift in the smoothed
signal.
• Long-term CD detects significant change within a pre-defined period that develops
through smaller drops in the margin.
Change detection - trending
Ambient conditions
Different data distributions
Comparison of sister engines
Time series – vector transpose
10
start_n2 start_egt
7560 209
8156 228
6858 102
8049 152
8396 208
8368 208
8031 165
6860 123
6982 122
6977 122
start_n2 start_egt
756
0 8156 6858 8049 8396 8368 8031 6860 6982 6977 209 228 102 152 208 208 165 123 122 122
score
10.7
11.6
8.4
10.3
11.5
11.5
10.5
8.7
8.8
8.8
score
14.5
SVM
SVM
Time series – vector transpose
11
start_n2 start_egt
756
0 8156 6858 8049 8396 8368 8031 6860 6982 6977 209 228 102 152 208 208 165 123 122 122
654
2 7053 7127 6981 6970 6985 7204 7126 8159 7954 188 206 223 169 154 183 209 212 143 158
655
7 7245 8641 8022 7684 7540 6837 7021 7489 7124 187 164 116 202 198 178 168 220 185 182
start_n2 start_egt
756
0 8156 6858 8049 8396 8368 8031 6860 6982 6977 209 228 102 152 208 208 165 123 122 122
815
6 6858 8049 8396 8368 8031 6860 6982 6977 6542 228 102 152 208 208 165 123 122 122 188
685
8 8049 8396 8368 8031 6860 6982 6977 6542 7053 102 152 208 208 165 123 122 122 188 206
Moving window
Separate rows
Feature binning
12
Moving window: 30
rec.
egt_corrected &
scaled
1 -0.0603 16 0.9502
2 -1.0557 17 -0.4373
3 0.3771 18 -0.4524
4 -0.2262 19 -0.4976
5 -0.3619 20 -0.754
6 -1.1763 21 -0.8144
7 0.1207 22 -0.5881
8 -1.1763 23 -0.2111
9 -1.0557 24 -1.1311
10 -1.3724 25 0.0001
11 -0.8596 26 -0.3921
12 -1.5534 27 -1.9153
13 0.166 28 0.8296
14 -0.8898 29 0.3319
15 0.6486 30 1.3273
Bin 1 Bin 2 Bin 3 Bin 4 Bin 5
3 12 9 4 2
Output
Support Vector Machine
13
Support Vector Machine
14
One Class SVM
15
Results of Start Monitoring
16
Rolldown differences: interesting
cases
17
18
Rolldown differences: interesting
cases
Damage detection
Difference between two signals
Difference between two signals
Difference between two signals