(AdaBN) with average accuracy being 95.95%. This result prove that the features learned by

A2CNN are more domain invariant than the features learned by the other methods.

training can significantly improve the bearing fault diagnosis under varying working conditions.

It is also interesting that when adapting from Domain A to B, from B to A, from B to C,

and from C to B, the fault diagnosis accuracy of the proposed A2CNN is only a bit better than

WDCNN (AdaBN). However, when adapting from domain A to C and C to A, the proposed

A2CNN is significantly better than the other methods. This result prove that A2CNN is good at

solving the problem that the distributions of source domain and target domain are far di

fferent.

4.4. Sensitivity Analysis of Faults

For each type of fault detection, in order to further analyze the sensitivity of the proposed

A2CNN model, we introduce two new evaluation indicators, i.e. precision and recall, which are

widely used in pattern recognition, information retrieval and binary classification.

In the fault diagnosis context, the precision and recall for a kind of fault type f can be

calculated as below,

precision

( f )

=

T P

T P

+ FP

, recall( f ) =

T P

T P

+ FN

,

(7)

where true positives (T P) means the number of faults correctly identified as f , false positives

(FP) means the number of faults incorrectly labeled as f and false negatives (FN) means the

number of faults f incorrectly labeled as not belonging to f .

A precision score of 1.0 for a fault type f means that every sample labeled as belonging

to class f does indeed belong to class f (i.e. there is no false alarm), but it can’t tell us about

the number of samples from class f that were not labeled correctly (i.e. how many failures are

missing?).

Whereas a recall of 1.0 means that every item from a fault type f was labeled as belonging

to class f (i.e. there is no missing alarm), but says nothing about how many other items were

incorrectly also labeled as belonging to class f (i.e. how many false alarms are there?).

Table 3 and Table 4.

precision

when adapting from domain B to C and from C to A, which are 49.63% and 57.55%

respectively. This means that about half of that kind of fault alarms are unreliable.

Meanwhile, in Table 4, for the 2nd kinds of fault (i.e. IF with fault size being 0.007 in.),

1.13% and 26.75% respectively. This means that about a large number of that kind of failures

are not detected.

plies that A2CNN has fewer false alarms (i.e. high precision score ) and missed alarms (i.e.

high recall score ). We can find that A2CNN can make almost all class classified into right

class, except BF with fault size being 0.014 in and BF with fault size being 0.021 in. This result

shows that after adversarial fine tuning, the classification performance on every class achieve

remarkable improvement.

12