Two-phase tweedie exponential dispersion process for degradation modeling: An adaptive Bayesian synthetic likelihood approach

Abstract

Engineered systems operate in complex and dynamic environments where their degradation mechanisms and failure modes may change. However, traditional degradation models, assuming the degradation of systems being a single pattern (or called single phase), cannot adequately capture the fashion that systems' degradation patterns varying from one phase to another. To fill this research gap, this study introduces a new two-phase degradation model based on the Tweedie exponential dispersion process (TEDP). The model employs the TEDP to characterize two phases of degradation behaviors while accounting for randomness in the change-point and degradation rates. The multiple filter algorithm (MFA) is introduced to detect the change-point. To tackle the complicated likelihood function arising from the two-phase TEDP, an adaptive Markov chain Monte Carlo Bayesian synthetic likelihood (MCMC-BSL) algorithm is developed to estimate the model parameters. The MCMC-BSL algorithm evaluates the likelihoods by simulating the model and analyzing summary statistics of data to secure both accuracy and computational efficiency of parameter estimation. The effectiveness of the proposed method is demonstrated through data numerical example and real lithium-ion batteries datasets.