3. During the stage of fine tuning, with the increasing of the adversarial iterations, the distri-

it

T

2000

T

and 99.90% respectively. The observation indicates that the adversarial network is e

ffec-

tive in improving the domain adaptation capacity of the A2CNN.

−100

−50

0

50

100

−60

−40

−20

0

20

40

60

80

(a)

−100

−50

0

50

100

−80

−60

−40

−20

0

20

40

60

80

(c)

NO

0.007/IF

0.014/IF

0.021/IF

0.007/BF

0.014/BF

0.021/F

0.007/OF

0.014/OF

0.021/OF

−100

−50

0

50

100

−80

−60

−40

−20

0

20

40

60

80

(b)

NO

0.007/IF

0.014/IF

0.021/IF

0.007/BF

0.014/BF

0.021/F

0.007/OF

0.014/OF

0.021/OF

Figure 4: Visualization of the extracted features of samples collected from the source domain B and target domain C via

t-SNE

S

}|

i

T

}|

i

by ten di

fferent colors respectively.

max outputs of the label predictor C. We randomly select a sample of the fault type ’0.007

/IF’

from domain C, denoted by x

0.007/IF

C

, as the input of the A2CNN model trained for adapting from

domain B to domain C. The visualized results are shown in Figure 5(a) and Figure 5(b).

From these visual results, we can find out that the output of the first three convolutional layers

(i.e. ’Conv1’, ’Conv2’ and ’Conv3’) are very similar. Starting from the fourth layer convolution

0.007/IF

C

0.007/IF

C

) gradually change

to some extent. This observation is consistent with the result in the section 4.5. That is, for the

domain adaptation from B to C, we only have to untie last four layers starting from ’Conv4’ to

’FC2’ as the features extracted from the the first three convolutional layers are almost the same.

The last fully-connected layer (i.e. ’FC2’) has K neurons (equals to the number of class

label). The output of ’FC2’ is fed to label classifier C to estimate the posterior probability of

each class using soft-max regression model.

According to the results of ’FC2’ and ’Softmax’ in Figure 5(a), x

0.007/IF

C

is misdiagnosed

as the fault type of ’0.014

.007/IF

C

). As a contrast,

in Figure 5(b), x

0.007/IF

C

is correctly identified as as the fault type of ’0.007

/IF’, based on the

0.007/IF

C

).

5. Conclusion

This paper proposes a adversarial adaptive model based on 1-D CNN named A2CNN, to

address the fault diagnosis problem under varying working condition. To our best knowledge,

15