A->B

A->C

B->A

B->C

C->A

C->B

AVG

SVM

68.60%

60.00%

73.20%

67.60%

68.40%

62.00%

66.63%

MLP

82.10%

85.60%

71.50%

82.40%

81.80%

79.00%

80.40%

DNN

82.20%

82.60%

72.30%

77.00%

76.90%

77.30%

78.05%

WDCNN

99.20%

91.00%

95.10%

91.50%

78.10%

85.10%

90.00%

WDCNN(AdaBN)

99.40%

93.40%

97.50%

97.20%

88.30%

99.90%

95.95%

A2CNNs

99.86%

98.40%

97.89%

89.46%

89.65%

99.14%

95.73%

A2CNN

99.99%

99.30%

98.18%

99.90%

97.93%

99.99%

99.21%

50.00%

60.00%

70.00%

80.00%

90.00%

100.00%

Figure 2: Accuracy (%) on six domain adaptation problems.

No.

Layer type

Kernel

stride

Channel

Output

Padding

1

Convolution1

32 × 1

2 × 1

8

1009 × 8

Yes

2

Pooling1

2 × 1

2 × 1

8

504 × 8

No

3

Convolution2

16 × 1

2 × 1

16

245 × 16

Yes

4

Pooling2

2 × 1

2 × 1

16

122 × 16

No

5

Convolution3

8 × 1

2 × 1

32

58 × 32

Yes

6

Pooling3

2 × 1

2 × 1

32

29 × 32

No

7

Convolution4

8 × 1

2 × 1

32

11 × 32

Yes

8

Pooling4

2 × 1

2 × 1

32

5 × 32

No

9

Convolution5

3 × 1

2 × 1

64

2 × 64

Yes

10

Pooling5

2 × 1

2 × 1

64

1 × 64

No

11

Fully-connected

500

1

500

12

Fully-connected

10

1

10

of adversarial adaptation, we use the corresponding output of the well-trained source feature

i

i

∈ {1, ..., K}

4.3. Accuracy across di

fferent domains

As Figure 2 shows, SVM, MLP and DNN perform poorly in domain adaptation, with average

accuracy in the six scenarios being around 65%, 80% and 80%. Which prove that samples under

di

fferent working conditions draw from the different distributions and models trained under one

working condition is not suitable for fault classification under another working load condition.

Compared with the WDCNN with AdaBN, which achieved the state-of-art domain adaptation

ability, A2CNN which achieves 99.21% accuracy in average is obviously greater than WDCNN

11