Detection of unknown bearing faults using re-weighted symplectic geometric node network characteristics and structure analysis

• Published in: Expert systems with applications 2023-04, Vol.215, p.119304, Article 119304
• Main Authors: Wang, Nini, Ma, Ping, Wang, Xiaorong, Wang, Cong, Zhang, Hongli
• Format: Article
• Language: English
• Subjects: Autocorrelation kurtosis; Community structure; Fault diagnosis; Re-weighted symplectic geometric node network; Unknown fault detection
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2022.119304

•Proposing re-weighted symplectic geometric node network based on complex network.•The Autocorrelation kurtosis and the ΓR are defined to eliminate the interference.•The CN theory was appiled to solve the misdiagnosis in unknown fault detection.•Using five groups of datasets to analysis the proposed method in various aspects.•Comparative analysis has been performed with the relevant advanced methods. When appears unknown fault type signal, typical classifiers would misclassify this fault to other types. And when equipment works, it will be accompanied by strong background noise, resulting in low accuracy of fault diagnosis. Therefore, a detection method based on re-weighted symplectic geometric node network (RSGNN) characteristics and structure analysis is proposed for noise reduction and unknown fault detection in this paper. First, the original signal was projected into network form to eliminate the interference components and characterize the inner topological relation. Then, the complex network (CN) theory was introduced to establish enhanced diagnosis features. Finally, the unsupervised Newman Fast (NF) community clustering algorithm was utilized to realize the unknown type signal detection. On the one hand, the proposed method defines Autocorrelation kurtosis (AK) and the ΓR to help better select the useful components and to eliminate the noise interference. On the other hand, the proposed method with the help of CN theory and community division to realize the unknown fault unsupervised detection. The experimental results on five groups bearing datasets show that the proposed method could achieve good unknown fault detection accuracy and maintain good robustness in a high-noise environment, indicating that it has advantages over other methods.