Speaker: Harish Kaushik, Tampere University
Reliability and Probabilistic Risk / Safety Assessment (PRA / PSA) of components in complex infrastructure systems is a major concern in disciplines including nuclear safety and civil engineering. In nuclear power plants (NPP) a prominent modelling strategy, Bayes-Empirical-Bayes (BEB), models failure rates of their components such as diesel generators, pumps, etc. arising from a mixture of two Poisson processes. However, there still remains challenges in failure rate estimation, especially when the data is zero-inflated. The objectives are to understand the development of the BEB model's hyperparameters, and to predict failure accurately using synthetic data. While variations of the BEB model and simple zero-inflated models have been investigated in this field, our novel focus is on reconstructing the failure distribution of each component using Markov Chain Monte Carlo (MCMC); and using novel hierarchical zero-inflated Bayesian models which have not been investigated in the domain of PRA within NPPs. In our work, we first investigate the consequences of choosing a particular type of scaling for the MCMC proposal kernel for BEB -- linear or logarithmic; and understand which sampling strategy works better for a particular type of data. To mitigate zero-inflation, we investigate two variants of the BEB likelihood -- Zero-Inflated Negative Binomial (ZINB) and Hurdle Negative Binomial (HNB). We characterise the ability of these models to recover ground truth component failure rates with simulated zero-inflated data in a comprehensive controlled setting. The primary motivation is to improve PRA in NPPs, however the Bayesian modelling framework and inference methods could be applied and transferred across other structural, and mechanical engineering domains especially for zero-inflated count data.
Where: Meeting room, 1st floor, IAC, Via dei Taurini 19
Organized by Annalisa Pascarella
