Kriging-Assisted Dynamic Bayesian Network for Assessing Lifecycle Risks of Lithium-Ion Batteries

Reliable lifecycle risk assessment of lithium-ion batteries (LIBs) is critical for safe, effective energy management in electric vehicles and energy storage systems, where unforeseen degradation or failure can trigger significant safety hazards and economic losses. In this study, we propose a Kriging-assisted Dynamic Bayesian Network (DBN) approach for dynamically predicting degradation processes of LIBs and quantifying time-dependent risk levels over the lifecycle. Within the DBN structure, LIB state variables, including State of Health (SOH) and Remaining Useful Life (RUL), are probabilistically coupled with degradation progression and thermal-runaway risk estimation. The interdependent dynamics generate temporal trajectories that are able to quantify lifecycle risk profiles of LIBs. The Kriging model is then employed into the DBN quantitative analysis to predict the mean and variance of child variables across continuous parent domains, enabling smooth probabilistic inference and uncertainty quantification. Conditional Probability Tables (CPTs) are constructed through Monte Carlo integration based on the Kriging-predicted mean to obtain continuous conditional probabilities, while State Transition Matrices (STMs) are jointly parameterized by both the mean and variance to capture temporal transition dynamics across time slices. This hybrid learning strategy decouples deterministic CPT estimation from stochastic STM propagation, integrates physics-informed priors, and prevents double-counting of uncertainty. Experimental validation on NASA battery cycling datasets and TUM venting-temperature data demonstrates that the proposed framework yields stable CPT and STM estimates, enables accurate inference of SOH and RUL, and produces physically consistent risk trajectories.