Back to search

Virginia Tech

Time-to-Event Modeling with Bayesian Perspectives and Applications in Reliability of Artificial Intelligence Systems

Abstract

dc:description.abstract

Time-to-event models are essential in assessing the reliability of materials, devices, and complex systems. With the rising need to analyze the safety of artificial intelligence (AI) systems, new, flexible methods in reliability analysis need to be developed. There are three types of data in reliability analysis: lifetime data, recurrent events data, and degradation data. This dissertation mainly focuses on statistical modeling of the first two data types using state-of-the-art reliability models. The proposed methods are motivated by two application data sets: the autonomous vehicles (AV) recurrent events data and the graphics processing unit (GPU) failure time data. In modeling, Bayesian analysis is commonly used to incorporate the prior knowledge of the model, and Markov chain Monte Carlo (MCMC) is a popular method for estimation and inference. This dissertation applies the Bayesian framework and No-U-Turn Sampler (NUTS) to analyze GPU failure time data. Motivated by the need to address the long computing time in using NUTS, a new MCMC sampler is then developed to sample from spatial survival models. In the dissertation, Chapter 1 provides a general introduction of the time-to-event data, the Bayesian framework, the reliability of hardware and software in AI systems, and the AV and GPU applications. Chapter 2 introduces the reliability analysis of AI systems in AVs. A statistical framework for modeling and analyzing the recurrent events data from AVs is developed. A new nonparametric model based on monotonic splines is proposed to estimate the cumulative baseline intensity function of the event process. The reliability of AVs from four manufacturers is evaluated using the proposed method, based on the data from California Department of Motor Vehicles. Chapter 3 studies the failure event time of GPUs in the supercomputer Titan for two different failure types. A model that jointly considers competing risks, spatial correlations among different GPU cabinets, and the presence of other covariates is proposed. In exploring the spatial correlation, the model accounts for both the physical location of server cabinets and the connectivity of cabinets. A logical circle distance is introduced in the spatial correlation function, and the non-positive definiteness issue of including circle distance is solved. A Bayesian framework is used for statistical inferences, and flat priors are added to most model parameters. Chapter 4 develops a block No-U-Turn Sampler (BNUTS) that is useful for sampling from spatially correlated accelerated failure time (SC-AFT) models. The proposed BNUTS is based on two popular samplers, the Gibbs sampler and NUTS. For the SC-AFT models with a large number of random effects, the proposed BNUTS can obtain samples with reasonably small auto-correlations within a shorter amount of time compared with NUTS. The factors that influence the computing time and the effective sample size of BNUTS are discussed. The new sampler is then applied to analyze the GPU failure time data, showing the capability of solving real-world research problems. Chapter 5 summarizes the main results of the dissertation and discusses several future research topics.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Statistics
Department dc:contributor.department
Statistics
Grantor dc:publisher
Virginia Tech
Year dc:date.issued
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Min, Jie
Chair dc:contributor.committeechair
  • Hong, Yili
Committee members dc:contributor.committeemember
  • Franck, Christopher Thomas
  • Kim, Inyoung
  • Datta, Jyotishka

Subjects

dc:subject × 6

Rights

dc:rights
Statement dc:rights
  • In Copyright
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Other
vt_gsexam:41140

Chain of custody

source
Harvested from
Virginia Tech
Base URL
vtechworks.lib.vt.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Min, Jie. Time-to-Event Modeling with Bayesian Perspectives and Applications in Reliability of Artificial Intelligence Systems. doctoral thesis, Virginia Tech, 2024. https://hdl.handle.net/10919/120579