A Prognostic Model Based on DBN and Diffusion Process for Degrading Bearing

• Published in: IEEE transactions on industrial electronics (1982) Vol. 67; no. 10; pp. 8767 - 8777
• Main Authors: Hu, Chang-Hua, Pei, Hong, Si, Xiao-Sheng, Du, Dang-Bo, Pang, Zhe-Nan, Wang, Xi
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.10.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Bearing degradation; Belief networks; deep belief network (DBN); Degradation; diffusion process (DP); Feature extraction; first hitting time (FHT); Indexes; locally linear embedding (LLE); Machine learning; Prediction models; Probability density functions; Reliability; remaining useful life (RUL); Signal monitoring; Uncertainty
• ISSN: 0278-0046; 1557-9948
• DOI: 10.1109/TIE.2019.2947839

Remaining useful life (RUL) prediction is extremely significant to ensure the safe and reliable operation for bearing suffering from the deterioration. The main focus of the RUL prediction is to accurately predict the future failure event, and thus, how to quantify the prediction uncertainty will be a major concern. However, current deep learning based RUL prediction methods are difficult to reflect the uncertainty of the RUL prediction results. Toward this end, we propose a RUL prediction model based on the deep belief network (DBN) and diffusion process (DP) in this article. The proposed method consists of two parts: feature extraction combining DBN and locally linear embedding (LLE), DP-based RUL prediction. In the first part, DBN is used to extract deep hidden features behind the monitoring signals, and then the features with higher tendency are screened as the input of LLE. The health index that can truly reflect the bearing health condition is further determined through LLE. In the second part, a health index evolving model based on DP is presented and the probability density function (PDF) of the predicted RUL is accordingly derived in the sense of the first hitting time (FHT). As such, the proposed method holds promise to improve the prediction accuracy and facilitate the prognostic uncertainty. Finally, experimental studies on the bearing degradation data and the associated comparative analysis verify the effectiveness and superiority of the proposed method.