An improved generative adversarial network for fault diagnosis of rotating machine in nuclear power plant

•A transfer learning method is proposed for fault diagnosis of rotating machinery under variable speed.•An adversarial training strategy driven by subdomain adaptation is proposed for classification problem.•Comparative cases of different detection methods, training efficiency, test results and feat...

Celý popis

Uloženo v:

Podrobná bibliografie
Vydáno v:	Annals of nuclear energy Ročník 180; s. 109434
Hlavní autoři:	Wang, Zhichao, Xia, Hong, Yin, Wenzhe, Yang, Bo
Médium:	Journal Article
Jazyk:	angličtina
Vydáno:	Elsevier Ltd 01.01.2023
Témata:	Fault diagnosis Generative adversarial network Nuclear power plant Objective function optimization Rotating machine Fault diagnosis Objective function optimization Rotating machine Nuclear power plant Generative adversarial network
ISSN:	0306-4549, 1873-2100
On-line přístup:	Získat plný text
Tagy:	Přidat tag Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!

Popis
Shrnutí:	•A transfer learning method is proposed for fault diagnosis of rotating machinery under variable speed.•An adversarial training strategy driven by subdomain adaptation is proposed for classification problem.•Comparative cases of different detection methods, training efficiency, test results and feature visualization are considered.•Relevant real-world experiments capture common rotating machines failures for data support. Due to the wide application and high safety requirements of rotating machines in nuclear power plants (NPPs), it has received more and more attention. The rotating machines fault diagnosis methods integrating vibration signal monitoring and deep learning provide reliable operation support solution for NPPs equipment. However, the inconsistent data probability distributions at different power levels limit the application of traditional models, which is a practical and neglected problem. To enhance the model generalization ability when the data are inconsistent, a deep transfer learning method based on hybrid domain adversarial learning (HDAL) strategy is proposed. First, the time–frequency domain features are extracted from the three-channel monitoring signals and fused into RGB images. Then, the deep convolutional network is applied to adaptively extract transferable features. More significantly, the transfer learning approach with an improved adversarial training strategy is implemented to more fine-grained reduce feature discrepancy, accordingly diagnosing faults under target power level. Finally, the general characteristics of typical faults are comprehensively captured based on rotating machines operation experiments, which provide data support for case testing. The comparative test and feature visualization verify the superiority of the proposed method, demonstrating the potential application value in NPPs rotating machines.
ISSN:	0306-4549 1873-2100
DOI:	10.1016/j.anucene.2022.109434