Failure-Mode-Informed Development of Remaining Useful Life Prognostics

PHM Society European Conference, vol. 9, pp. 1–13

Abstract

Various environmental and operating conditions affect the degradation behavior of physical assets, leading to various degradation trajectories and ultimately to distinct failure modes. To obtain accurate Remaining Useful Life (RUL) prediction, it is important to distinguish between such degradation trajectories and their associated failure modes. In this paper, we develop a framework where we analyze the latent space of autoencoders using spectral clustering to evaluate the similarity in degradation trajectories and failure modes in training datasets. This failure-mode-informed training sets are then used to develop failure-specific regressors for RUL prediction. On one hand, this reduces the amount of data needed to effectively train prognostics models. In addition, the accuracy of the RUL predictions is further improved. We demonstrate this using the C-MAPSS dataset, which provides fleet-based run-to-failure sequences under varying operating conditions and failure modes. We argue that latent information about different degradation mechanisms can be inferred from sensor readings, enabling the construction of failure-mode-specific RUL regressors. Our results show that this failure-mode-informed data separation reduces the amount of training data needed to generate RUL prognostics by up to 55%, while simultaneously improving prognostics accuracy - the Root Mean Square Error (RMSE) is reduced by 3%.

Authors 4

1. Kiavash Fathi

   ZHAW Zurich University of Applied Sciences

   Affiliation as printed

   Institute of Mechatronic Systems, Zurich University of Applied Sciences, 8400 Winterthur, Switzerland

2. Mihaela Mitici

   Utrecht University

   Affiliation as printed

   Faculty of Science, Utrecht University, Heidelberglaan 8, 3584 CS, Utrecht, The Netherlands

3. Tobias Kleinert Aachen Chair of Information and Automation Systems for Process and Material Technology

   RWTH Aachen University

   Affiliation as printed

   Chair of Information and Automation Systems for Process and Material Technology, RWTH Aachen University, 52064 Aachen, Germany

4. Hans Wernher van de Venn

   ZHAW Zurich University of Applied Sciences

   Affiliation as printed

   Institute of Mechatronic Systems, Zurich University of Applied Sciences, 8400 Winterthur, Switzerland