Table 1: Description of the CWRU dataset

Category labels

1

2

3

4

5

6

7

8

9

10

Fault location

None

IF

BF

OF

Load

Fault diameter (in.)

0

0.007

0.014

0.021

0.007

0.014

0.021

0.007

0.014

0.021

Domain A

800

800

800

800

800

800

800

800

800

800

1

Domain B

800

800

800

800

800

800

800

800

800

800

2

Domain C

800

800

800

800

800

800

800

800

800

800

3

#Features

2048

2048

2048

2048

2048

2048

2048

2048

2048

2048

To construct domain adaptation problems, we randomly choose two from the three domains,

where one is considered as the source domain and the other is considered as the target domain.

A → B

4.2. Experimental setup

4.2.1. Baseline Methods

We compare our methods with the following baselines,

1. Traditional SVM and Multi-layer Perceptron (MLP) which work with the data transformed

by Fast Fourier transformation (FFT).

2. The deep neural network (DNN) system with frequency features [10] proposed by Lei et

al. in 2016. This neural network consists of three hidden layers. The number of neurons

in each layer is 1025, 500, 200, 100 and 10. The input of the network is the normalized

1025 Fourier coe

fficients transformed from the raw temporal signals using FFT.

3. The Deep Convolution Neural Networks with Wide first-layer kernels (WDCNN) system

[26] proposed by Zhang et al. in 2017. The WDCNN system works directly on raw

temporal signals. It contains five convolutional layers and batch normalization layers.

The domain adaptation capacity of this model originates in the domain adaptation method

named Adaptive Batch Normalization (AdaBN).

4.2.2. Parameters of the proposed A2CNN

layers and pooling layers followed by two fully-connected hidden layers. The pooling type

is max pooling and the activation function is ReLU. The parameters of the convolutional and

pooling layers are detailed in Table 2. In order to minimize the loss function, the Adam Stochastic

optimization algorithm is applied to train our CNN model. The final l (l ∈ [1, 7]) layers of the

consists of three fully-connected layers. The number of neurons in each layer is 500, 500 and 1.

The experiments were implemented using Tensorflow toolbox of Google.

10