A graph neural network and Bayesian inference–based method for reliability assessment of special vehicles under small-sample and zero-failure conditions

Abstract Special vehicles are typical complex systems. They feature diverse functions, strong structural coupling, and high reliability requirements. These properties make reliability methods that rely on abundant failure data unsuitable. To address low accuracy under small-sample or zero-failure conditions, this study proposes a vehicle reliability assessment framework that integrates graph neural networks and Bayesian inference. First, a duallayer heterogeneous graph is constructed for the vehicle system and multi-phase tests. It captures component topology and test-stage evolution. Next, a graph-guided adaptive Bayesian inference scheme is developed. Graph neural networks dynamically construct prior distributions. A graph-guided MCMC algorithm improves sampling efficiency and convergence in high-dimensional posteriors. A graph-propagation-based structured inference strategy is further introduced to enable robust reliability estimation in zero-failure scenarios. The results show that the proposed method effectively combines system structure and multi-source test data, improving parameter estimation accuracy under small samples. In particular, under zero-failure conditions, the graph-propagation strategy leverages neighbor information to produce reasonable reliability estimates. This provides a feasible solution for special vehicle reliability assessment with small samples or no failures.