BibTeX

@InProceedings{Lewis2019ESREL,
author = {Lewis, Austin D. and Groth, Katrina M.},
title = {A Review of Methods for Discretizing Continuous-Time Accident Sequences},
booktitle = {Proceedings of the 29th European Safety and Reliability Conference {(ESREL 19)}},
year = {2019},
eventdate = {2019-09-22/2019-09-26},
abstract = {Operators and engineers rely on accident sequence data from complex systems to develop a greater understanding of their failure points. Certain systems, such as nuclear reactors, have limited accident data from actual failures; in these systems, models are used to produce simulated accident data and training scenarios to supplement the available data. These simulations are designed to efficiently create accurate and meaningful system data that can provide the operators with diagnostic and prognostic capabilities.
Prognostics models can reduce computational costs by representing a continuous-time accident sequence as a discrete-time model, such as a Dynamic Bayesian Network (DBN). However, there are open questions about how operational time series data can and should be discretized to maximize diagnostic and prognostic capabilities. This paper examines different approaches that could be used to partition the time sequence into discrete-time slices. A survey of the reliability engineering literature indicates that researchers primarily select one of two system-independent methods for generating time slices depending on the nature of the study; however, other methods can take advantage of the dynamic changes in system parameter states to provide a clearer picture of major system features. Combining both approaches in a model would generate meaningful system accident data that could increase system diagnostic and prognostic capabilities while also limiting the model’s computational burden.
},
address = {Hannover, Germany},
file = {:Conference Papers/Lewis_ESREL_2019_DiscretizationTechniques.doc:Word},
keywords = {Dynamic Bayesian Network; time series; time discretization; accident sequence simulation},
timestamp = {2019-05-17},
}

Abstract

Operators and engineers rely on accident sequence data from complex systems to develop a greater understanding of their failure points. Certain systems, such as nuclear reactors, have limited accident data from actual failures; in these systems, models are used to produce simulated accident data and training scenarios to supplement the available data. These simulations are designed to efficiently create accurate and meaningful system data that can provide the operators with diagnostic and prognostic capabilities.
Prognostics models can reduce computational costs by representing a continuous-time accident sequence as a discrete-time model, such as a Dynamic Bayesian Network (DBN). However, there are open questions about how operational time series data can and should be discretized to maximize diagnostic and prognostic capabilities. This paper examines different approaches that could be used to partition the time sequence into discrete-time slices. A survey of the reliability engineering literature indicates that researchers primarily select one of two system-independent methods for generating time slices depending on the nature of the study; however, other methods can take advantage of the dynamic changes in system parameter states to provide a clearer picture of major system features. Combining both approaches in a model would generate meaningful system accident data that could increase system diagnostic and prognostic capabilities while also limiting the model’s computational burden.