Speaker: 徐安察教授，浙江工商大学

Inviter: 胡庆培研究员

Title: Adaptive Heterogeneity Learning in Nonlinear Wiener Degradation Models via Dirichlet Process Mixtures

Language: Chinese

Time & Venue: 2026.04.24 15:30-16:30 南楼N202

Abstract: Degradation-based reliability analysis heavily relies on random effects to capture unit-to-unit heterogeneity. However, standard approaches impose a parametric form on these latent variables, incurring significant misspecification risk when the true population exhibits multimodality or skewness. We propose a robust semiparametric Bayesian framework for generalized nonlinear Wiener processes in which the drift-based random effects follow a Dirichlet process mixture, letting the data adaptively determine the complexity of the population structure. We derive analytical approximations to the first-passage-time distribution that retain closed-form expressions for the reliability function and mean-time-to-failure. Posterior inference is performed via an efficient hybrid blocked Gibbs sampler coupled with slice sampling. Extensive simulations show that the method consistently recovers the true latent densities and yields accurate lifetime predictions even under severe distributional misspecification. Two real datasets demonstrate the framework’s ability to uncover latent sub-populations and tighten predictive intervals.