An artificial neural network supported stochastic process for degradation modeling and prediction. (October 2021)

• Record Type: Journal Article
• Authors: Liu, Di; Wang, Shaoping
• Is Part Of: Reliability engineering & system safety. Volume 214(2021)
• Publication Date: 2021-10
• Subjects: Degradation modeling -- Degradation prediction -- Stochastic process -- Degradation path uncertainty -- Artificial neural network; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2021.107738
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: Stochastic process is combined with ANN to handle degradation path uncertainty. The hyper parameters are evaluated by moment estimation offline. The process parameters are updated by Bayesian inference for online prediction. Without path information the ANN supported stochastic process is still practical. The assumption that the initial degradation is zero is also freed. Abstract: An artificial neural network supported stochastic process for degradation modeling and prediction is proposed in this paper. An artificial neural network is applied to describe the degradation path due to data fitting flexibility and path description considering degradation path uncertainty. The assumption that the initial degradation is zero in the stochastic process is freed. The artificial neural network supported stochastic process is trained by minimizing the minus log-likelihood offline-based on the run-to-failure degradation data. Considering unit-to-unit variance in population degradation modeling, the process parameters are assumed to be randomly distributed. Here, three common distributions describe the process parameters, and Akaike information criteria applied to select distributions of process parameters. The process parameters are evaluated by Bayesian inference based on the trained path, and distributions of process parameters are based on real-time degradation data for online prediction. The proposed method is verified by a simulation experiment based on a Wiener process, which seemed a true model. Furthermore, an actual degradation dataset is also applied to illustrate the effectiveness of the proposed method. Both the simulation experiment and actual example indicate that the proposed stochastic process is capable of degradation modeling and degradation predicting, even without prior information about the degradation path.