A2CNN datasheet(13/19 Pages) ETC1

13 / 19 page

Table 3: precision of the proposed A2CNNS and A2CNN on six domain adaptation problems.

Fault location

None

Fault diameter (in.)

0.007

0.014

0.021

0.007

0.014

0.021

0.007

0.014

0.021

Category labels

precision

of A2CNNS

A→B

100%

98.64%

100%

A→C

92.27%

100%

99.75%

92.82%

99.46%

100%

B→A

100%

87.43%

93.68%

100%

B→C

96.74%

100%

49.63%

100%

99.49%

100%

C→A

100%

57.55%

100%

83.33%

95.40%

100%

95.12%

100%

C→B

100%

93.13%

99.88%

98.89%

100%

precision

of A2CNN

A→B

100%

99.88%

100%

A→C

93.46%

100%

B→A

100%

90.70%

92.59%

100%

B→C

100%

99.01%

100%

C→A

100%

90.91%

90.40%

100%

C→B

100%

99.88%

100%

Table 4: recall of the proposed A2CNNS and A2CNN on six domain adaptation problems.

Fault location

None

Fault diameter (in.)

0.007

0.014

0.021

0.007

0.014

0.021

0.007

0.014

0.021

Category labels

recall

of A2CNNS

A→B

100%

98.63%

100%

A→C

100%

99.75%

100%

99.50%

92.13%

92.88%

100%

99.50%

100%

B→A

100%

78.88%

100%

B→C

100%

1.13%

100%

96.63%

99.50%

100%

97.38%

100%

C→A

100%

26.75%

100%

96.00%

73.75%

100%

C→B

100%

92.63%

100%

99.88%

98.88%

100%

recall

of A2CNN

A→B

100%

99.88%

100%

A→C

100%

93.00%

100%

B→A

100%

81.75%

100%

B→C

100%

99.00%

100%

C→A

100%

79.38%

100%

C→B

100%

99.88%

100%

4.5. Parameter Sensitivity

In this section, we investigate the inﬂuence of the parameter l, which represents the number

of untied layers in the target feature extractor MT during the adversarial adaptive ﬁne tune. Given

that the target feature extractor MT contains ﬁve convolutional layers and pooling layers and two

fully-connected hidden layers, l is selected from {1, ..., 7} in our experiment. We use A2CNNl to

denote the A2CNN model with the parameter l. For example, A2CNN1 indicates that only the

last fully-connected hidden layer is untied (i.e. FC2 in Figure 1), and A2CNN7 means all the

seven layers in MT are untied (i.e. from ’Conv1’ to ’FC2’ in Figure 1).

Figure 3 reports the results. From the ﬁgure, we can generally observe that the more untied

layers involved in adversarial adaptive ﬁne tuning stage, the higher the accuracy of recognition.

However, the sensitivity of di

ﬀerent adaptive problems to parameters l is diﬀerent.

First of all, when adapting from domain A to B and from B to A, the enhancement of recog-

nition accuracy is limited. We can use A2CNNS directly to achieve the accuracy of 99.86% and

97.89% respectively, which is only a little worse than A2CNN7. Moreover, in these two cases,

A2CNN Datasheet(PDF) 13 Page - List of Unclassifed Manufacturers

A2CNN Datasheet(HTML) 13 Page - List of Unclassifed Manufacturers

Html Pages

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Datasheet Download

Link URL