Abstract In recent years, neural network metamodels have become increasingly popular for reducing the computational burden of performing direct, simulation-based analysis of physical systems. This has led to a surge of interest in using neural networks and other machine learning tools to analyze and assess risk. This paper proposes a new methodology for training a neural network metamodel and incorporating it into a Bayesian network-based probabilistic risk assessment. This methodology can be applied to a wide variety of industrial accidents, where there is at least one latent variable that is normally calculated using a physics code. The main benefit of this methodology is that it combines the interpretability and sampling algorithm of a Bayesian network with the high-dimensional, latent variable modeling capability of a neural network metamodel. This paper also provides an example of how this methodology is applied to fissionable material operations in a nuclear facility, in order to estimate process criticality accident risk. Although process criticality accidents are specific to the nuclear industry, the methodology and generalized risk model described in this paper can be adapted to other types of industrial accidents and rare events.
Paper
Full text
Analysis of process criticality accident risk using a metamodel-driven Bayesian network
Semantic Scholar · Engineering · 2021
Abstract
Abstract In recent years, neural network metamodels have become increasingly popular for reducing the computational burden of performing direct, simulation-based analysis of physical systems. This has led to a surge of interest in using neural networks and other machine learning tools to analyze and assess risk. This paper proposes a new methodology for training a neural network metamodel and incorporating it into a Bayesian network-based probabilistic risk assessment. This methodology can be applied to a wide variety of industrial accidents, where there is at least one latent variable that is normally calculated using a physics code. The main benefit of this methodology is that it combines the interpretability and sampling algorithm of a Bayesian network with the high-dimensional, latent variable modeling capability of a neural network metamodel. This paper also provides an example of how this methodology is applied to fissionable material operations in a nuclear facility, in order to estimate process criticality accident risk. Although process criticality accidents are specific to the nuclear industry, the methodology and generalized risk model described in this paper can be adapted to other types of industrial accidents and rare events.