A new intelligent fault identification method based on transfer locality preserving projection for actual diagnosis scenario of rotating machinery. (1st January 2020)

• Record Type: Journal Article
• Authors: Zheng, Huailiang; Wang, Rixin; Yin, Jiancheng; Li, Yuqing; Lu, Haiqing; Xu, Minqiang
• Is Part Of: Mechanical systems and signal processing. Volume 135(2019)
• Publication Date: 2020-01-01
• Subjects: Transfer learning -- Intelligent fault diagnosis -- Locality preserving projection -- Rotating machinery; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2019.106344
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A new intelligent transfer diagnosis method of rotating machinery, TLPPIFI, is proposed. TLPPIFI can build the diagnosis model using historical data from other same-type machines. TLPPIFI performs superior transfer capacity with inadequate information of target domains. Three case studies show the validity and superiority of TLPPIFI. Abstract: Intelligent fault diagnosis methods have been widely developed in recent years due to the ability in learning diagnosis knowledge from monitoring data automatically. However, for many diagnosis methods based on traditional machine learning algorithms, how to collect massive data under the same distribution with test data is a difficult problem in real world industrial applications. Aiming at this data dilemma of conventional intelligent diagnosis methods, this paper proposes a T ransfer L ocality P reserving P rojection based I ntelligent F ault I dentification (TLPPIFI) method, which can construct diagnosis model using historical data collected from different operating conditions or other same-type machines. Based on a relevance assumption, TLPPIFI first embeds the data to a subspace through preserving a priori distribution structure properties of training data and minimizing the distribution discrepancy between different datasets simultaneously. By this means, the samples with same category in different datasets could cluster together in the new space. Finally, a classifier is trained to identify the condition of target machine by the historical data and the normal data of target machine together. The effectiveness of the proposed method is validated by three real-life diagnosis cases. The experimental results demonstrate that TLPPIFI can achieve superior diagnosis performance than several supervised learning methods and transfer learning methods. In addition, empirical analysis about distribution distance between domains and parameter sensitivity are also investigated.