3.2. Model Learning

We have used two training steps to enhance the domain adaptation ability of our model. The

details of the proposed algorithm is summarized in Algorithm 1.

2. Adversarial adaptive fine tune.Initialize the parameters of the target feature extractor

such that a domain discriminator D can not predict the domain label of mapped source and

target examples reliably.

3.3. Classifier Construction

After all the parameters are learned, we can construct a classifier for the target domain by

directly using the output of the last fully connected layer (i.e. ’FC2’) of the target feature ex-

i

i

i

∈ {1, ..., K}

using Eq. 2. We choose the maximum probability using Eq. 6. and the corresponding label as

the prediction,

y

i

j

K

l

, with u

j

(6)

4. Experimental analysis of proposed A2CNN model

In real world applications, data under di

fferent load condition usually draw from different

distribution. So it is significant to use unlabeled data under any load condition to rebuilt the

classifier trained with samples collected in one load condition. In the reminder of this section,

Case Western Reserve University (CWRU) bearing database is used to investigate how well the

proposed A2CNN method performs under this scenario.

4.1. Datasets and Preprocessing

The test-bed in CWRU Bearing Data Center is composed of a driving motor, a two hp mo-

tor for loading, a torque sensor

/encoder, a power meter, accelerometers and electronic control

unit. The test bearings locate in the motor shaft. Subjected to electro-sparking, inner-race faults

(IF), outer-race faults (OF) and ball fault (BF) with di

fferent sizes (0.007in, 0.014in, 0.021in and

0.028in) are introduced into the drive-end bearing of motor. The vibration signals are sampled

by the accelerometers attached to the rack with magnetic bases under the sampling frequency

of 12kHz. The experimental scheme simulates three working conditions with di

fferent mo-

tor load and rotating speed, i.e., Load1

= 1hp/1772rpm, Load2 = 2hp/1750rpm and Load3 =

3hp

/1730rpm. The vibration signals of normal bearings (NO) under each working condition are

also gathered.

In this paper, a vibration signal with length 4096 is randomly selected from raw vibration

signal. Then, fast Fourier transform (FFT) is implemented on each signal and the 4096 Fourier

coe

fficients are generated. Since the coefficients are symmetric, the first 2048 coefficients are

used in each sample. The samples collected from the above three di

fferent conditions form three

domains, namely A, B and C, respectively. There are ten classes under each working condition,

including nine kinds of faults and a normal state, and each class consists of 800 samples. There-

fore, each domain contains 8000 samples of ten classes collected from corresponding working

condition. The statistics of all domains are described in Table 1.

8