Collaborative deep learning framework for fault diagnosis in distributed complex systems. (July 2021)

• Record Type: Journal Article
• Authors: Wang, Haoxiang; Liu, Chao; Jiang, Dongxiang; Jiang, Zhanhong
• Is Part Of: Mechanical systems and signal processing. Volume 156(2021)
• Publication Date: 2021-07
• Subjects: Fault diagnosis -- Distributed complex systems -- Collaborative deep learning -- Privacy preserving; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.107650
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: Collaborative learning framework for fault diagnosis in distributed energy systems. Learning strategy utilizing local information with no need of raw data sharing. Consensus for distributed deep learning models which can be geographically located. The proposed framework outperforms the model trained with local data. The algorithm achieves close accuracies with central learning using multi-sensor data. Abstract: In distributed complex systems, condition monitoring and fault diagnosis have received considerable attention, especially for recent developments of data-driven methods with deep learning structures, which greatly enhance the performance as of superior representation capacity over big data. To apply these methods, massive data needs to be collected from distributed systems, requiring high costs for data transmission and causing more and more concerns on privacy issues. For the naturally distributed data in such scenario, this work presents a novel collaborative deep learning framework with the idea that the features, as representations of data, can be transmitted through latent parameters of deep learning structure while the raw data won't be shared in the distributed network. Based on the collaborative learning setup, the proposed framework adopts a secure communicating strategy with no need of transmitting raw data, and obtains a consensus for distributed deep learning models that can be geographically located. To validate the proposed scheme, four case studies are carried out and the results show that it is able to improve the diagnosis accuracy compared with local learning models. Also, it is robust and adaptive for diagnosis problems with data that is imbalanced or from different distributions.