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

adversarial network.

A2CNN contains four parts, a source feature extractor, a target feature extractor, a label clas-

sifier and a domain discriminator. In order to get the strong fault-discriminative and domain-

invariant capacity, we adopt the training process similar to Generating Adversarial Network

(GAN). First, in order to get fault-discriminative features, the source feature extractor is pre-

trained with labeled source training examples to minimize the label classifier error. Then, during

the adversarial adaptive fine tuning stage, the target feature extractor is initialized and trained to

maximize the domain classification loss, such that the domain discriminator can’t predict the do-

main label of mapped source and target examples reliably. That is to say, instances sampled from

the source and target domains have similar distributions after mapping. In addition, the layers

between the source and target feature extractor during the training stage are partially untied to

give consideration to the training e

fficiency and the domain adaptability.

Results in Section 4 shows that, compared with the state-of-the-art domain adaptive model

WDCNN (AdaBN), the proposed A2CNN achieve higher accuracy under di

fferent working con-

ditions. Besides the commonly used fault diagnostic accuracy, we introduce two new evaluation

indicators, precision and recall, to analyze the sensitivity of the proposed for each type of fault

detection. A precision score of 1.0 for a fault type means that there is no false alarm. Whereas a

recall

of 1.0 means that there is no missing alarm. Compared with accuracy, precision and recall

can evaluate the reliability of a model for certain type of fault recognition in more detail. This

result shows that after adversarial training, the classification performance on every class achieve

remarkable improvement. We can find that A2CNN can make almost every fault classified into

the right class. The high precision and recall score of A2CNN implies that our model has fewer

false alarms and missed alarms.

Finally, through visualizing the feature maps learned by our model, we explore the inner

mechanism of proposed model in fault diagnosis and domain adaptation, and verify that partially

untied the layers between the source and target feature extractor is correct.

References

References

[1] G. J. Vachtsevanos, F. Lewis, A. Hess, B. Wu, Intelligent fault diagnosis and prognosis for engineering systems,

Wiley Online Library, 2006.

[2] Z. Qiao, Y. Lei, J. Lin, F. Jia, An adaptive unsaturated bistable stochastic resonance method and its application in

mechanical fault diagnosis, Mechanical Systems and Signal Processing 84 (2017) 731–746.

[3] P. Albrecht, J. Appiarius, E. Cornell, D. Houghtaling, R. McCoy, E. Owen, D. Sharma, Assessment of the reliability

of motors in utility applications, IEEE Transactions on Energy Conversion EC-2 (3) (1987) 396–406. doi:10.

1109/TEC.1987.4765865.

[4] A. K. S. Jardine, D. Lin, D. Banjevic, A review on machinery diagnostics and prognostics implementing condition-

based maintenance, MECHANICAL SYSTEMS AND SIGNAL PROCESSING 20 (7) (2006) 1483–1510. doi:

{10.1016/j.ymssp.2005.09.012}.

[5] L. Ya-Guo, H. Zheng-Jia, Advances in applications of hybrid intelligent fault diagnosis and prognosis technique

[j], Journal of Vibration and Shock 9 (2011) 030.

[6] K. He, G. Gkioxari, P. Doll´ar, R. Girshick, Mask r-cnn, in: Computer Vision (ICCV), 2017 IEEE International

Conference on, IEEE, 2017, pp. 2980–2988.

[7] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille, Deeplab: Semantic image segmentation with

deep convolutional nets, atrous convolution, and fully connected crfs, IEEE transactions on pattern analysis and

machine intelligence 40 (4) (2018) 834–848.

17