A2CNN datasheet(6/19 Pages) ETC1

6 / 19 page

Pool1

��

Conv3

Conv2

Pool2

Pool3

Conv4

Pool4

Conv5

Pool5

Source Domain Examples + Labels

��

Target Domain Examples

Pool1

��

Conv3

Conv2

Pool2

Pool3

Conv4

Pool4

Conv5

Pool5

��,

��-

(Soft-max Regression Model)

Source Feature Extractor

Target Feature Extractor

��, ��/0 , 0|056

��- ��-0 , 1|056

��, ��/0 , ��,0|056

Domain Discriminator

��

Tied Layers

Adaptation Layers

Label Classifier

��

Figure 1: The proposed Adversarial Adaptive 1-D CNN (A2CNN) includes a source feature extractor MS , a target feature

extractor MT , a label classiﬁer C and a domain discriminator. Solid lines indicate tied layers, and Dashed lines indicate

adaptive layers.

Source feature extractor MS . As shown in Figure 1, we compose the source feature extractor

MS from ﬁve 1-D convolutional layers and two fully-connected layers. The input of the ﬁrst

convolution layer (i.e. ’Conv1’) is the fast Fourier transform (FFT) spectrum amplitudes of

vibration signals, which is the most widely used approach of bearing defect detection. The last

fully-connected layer (i.e. ’FC2’) is called label layer [37] with an output of K neurons (equals

to the number of class label), which is fed to label classiﬁer C which estimate the posterior

probability of each class. It is common to add a pooling layer after each convolution layer in

the CNN architecture separately. It functions as a down-sampling operation which results in

a reduced-resolution output feature map, which is robust to small variations in the location of

features in the previous layer. The most commonly used pooling layer is max-pooling layer,

which performs the local max operation over the input features. The main di

ﬀerence between

the traditional 2-D and the 1-D CNN is the usage of 1-D arrays instead of 2-D matrices for both

feature maps and ﬁlter kernels. In order to capture the useful information in the intermediate

and low frequency bands, the wide kernels should be used in the ﬁrst convolutional layer which

can better suppress high frequency noise[26]. The following convolutional kernels are small

(speciﬁcally, 3 × 1) which make the networks deeper to acquire good representations of the input

signals and improve the performance of the network.

Label classiﬁer C. For an source domain instance xi

S , the output feature vector MS (x

S )

∈ RK×1

mapped by the source feature extractor MS is the input of the label classiﬁer C. Here, the soft-

max regression model [38] is used as the label classiﬁer on source domain to incorporate label

information. The soft-max regression model is a generalization of the logistic regression model

for multi-class classiﬁcation problems. We can estimate the probabilities of each class that xi

A2CNN Datasheet(PDF) 6 Page - List of Unclassifed Manufacturers

A2CNN Datasheet(HTML) 6 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