taneously train two models: a generative model G captures the data distribution and a discrimi-

native model D estimates the probability that a sample came from the training data o generated

by G. Inspired by GAN, we designed a adversarial adaptive model based on 1-D CNN named

A2CNN, which simultaneously satisfied the above fault-discriminative and domain-invariant re-

quirements. The details of the model will be shown in Section 3. To our best knowledge, this

is the first attempt for solving the domain adaptation issues in fault diagnosis by introducing

adversarial network.

The main contributions of this literature are summarized as follows.

1) We propose a novel adversarial adaptive CNN model which consists of four parts, namely,

a source feature extractor, a target feature extractor, a label classifier and a domain discriminator.

During the training stage, the layers between the source and target feature extractor are partially

untied to take both training e

fficiency and domain adaptation into consideration.

2) This proposed model has strong fault-discriminative and domain-invariant capacity, and

therefore can achieve high accuracy under di

fferent working conditions. We visualize the fea-

ture maps learned by our model to explore the intrinsic mechanism of proposed model in fault

diagnosis and domain adaptation.

3) Besides the commonly used fault diagnostic accuracy, we introduce two new evaluation

indicators, precision and recall. Compared with accuracy, precision and recall can evaluate the

reliability of a model for certain type of fault recognition in more detail.

The rest of paper is organized as follows. In Section 2, some preliminary knowledge that will

be used in our proposed framework is briefly reviewed. Section 3 introduced the construction of

our proposed A2CNN. A series of experiments are conducted in Section 4. Finally, we conclude

this paper in Section 5.

2. Preliminary Knowledge

The above CNNs for fault diagnosis mentioned in 1 work well only under a common as-

sumption: The training and test data is drawn from the same distribution. However, vibration

signals used for fault diagnosis usually show disobedience of the above assumption. In the run-

ning process of rotating machinery, because of complicated working conditions, the distributions

of fault data under varying working condition are not consistent. For example, the training sam-

ples for building the classifier might be collected under the work condition without the motor

load, nevertheless the actual application is to classify the defects from a bearing system under

di

fferent motor load states. Although the categories of defects remain unchanged, the target data

distribution changes with the motor load varies.

Our ultimate goal is to be able to predict labels given a sample from one working condition

while the classifier is trained by the samples collected in another working condition. Then,

the problem above can be regarded as a domain adaptation problem, which is a realistic and

challenging problem in fault diagnosis. To solve this challenge, a domain adaption technique,

would be needed to learn a discriminative classifier or other predictor in the presence of a ”shift”

between training and test distribution by taking full advantage of information coming from both

source and target domains.

2.1. Domain Adaptation

According to the survey on domain adaptation (DA) for classification [29], a domain D

3