PSI - Issue 17

Takanori Hasegawa et al. / Procedia Structural Integrity 17 (2019) 487–494 Takanori Hasegawa et al./ Structural Integrity Procedia 00 (2019) 000 – 000

491

5

3.2 Experiment

Experimental comparisons using three-channel vibration data obtained from the actual FESD were conducted to investigate the effectiveness of the developed data-driven classifier on the operational condition monitoring of FESD.

3.2.1 Experimental result

Recorded data include four vibration signal sequences for each operation state. These sequences were divided and exploited for a four-fold cross validation test; for each validation, one sequence was used for testing and the remaining three sequences were used for training. Each sequence was divided every one second. Table 2 lists the number of one second data samples for each validation set and operation state. For each validation, the number of data samples for each state was adjusted to be equal for all validation sets. Note that training and testing data must not include same recording. Table 3 lists the confusion matrix in terms of the number of data samples identified into normal, no-helium, and bad-oil states for respective validation sets. This result indicates that the bad-oil state, which corresponds to an end stage of the FESD life, can be accurately identified whereas detecting no-helium state, which is a sign of failure, is difficult. In fact, the developed system confused the normal and no-helium states for the validation sets 1 and 4. This table also shows that classification accuracy varies according to validation sets. It implies that the mismatch in distributions between training and testing data cannot be negligible and an increase in training data is needed.

Table 2. Number of one-second data samples for each validation set. validation normal no-helium

bad-oil

train

test

train

test

train 5430 5430 5430 5430

test

15204 15204 15204 15204 60816

5068 5068 5068 5068

18462 18462 18462 18462 73848

6154 6154 6154 6154

1810 1810 1810 1810 7240

set 1 set 2 set 3 set 4 total

20272

24616

21720

Table 3. Performance of operation state classification. Numbers of data samples identified to normal, no-helium, and bad-oil conditions are listed.

Predicted label normal

Validation set

no-helium

bad-oil

5060

normal (set 1)

6154

0 0

0 0 0

no-helium (set 1)

8 0

1810

bad-oil (set 1) normal (set 2)

5037

0

6154

no-helium (set 2)

31

112

1698

bad-oil (set 2) normal (set 3)

0

0 1

5068

0

6153

no-helium (set 3)

0 0

486

1324

bad-oil (set 3) normal (set 4)

0 0

2282 2786

0

6065

no-helium (set 4)

142

1668

bad-oil (set 4)

0

89

3.2.2 Analysis

2500-dimensional spectra were projected onto two-dimensional space by principle component analysis (PCA) as shown in Fig. 4. This figure indicates that vibration data samples in the normal state (green symbols) are compactly