System reliability analysis based on non-parametric modeling and melding of multi-source data

In system reliability analysis, multi-source lifetime, degradation, and pass/fail data of multiple levels can compensate for the scarce data from full-system tests. However, existing research mainly uses parametric methods in modeling these data, which rely on strong prior assumptions on data uncertainty and correlation structures, thus having limited adaptability. To analyze system reliability more adaptively, a non-parametric methodology is developed. It begins with non-parametric modeling of single-source data, where adaptive kernel density estimation methods are proposed with adaptive bandwidth and local correlation coefficient to flexibly depict data uncertainty and correlation, respectively. They are verified to surpass commonly used parametric methods in modeling lifetime data with single and double peaks, and degradation data with linear, nonlinear, and double-cluster features. Next, rules for cross-level data transmission, multi-prior combination, and Bayesian melding based on equivalent parameterization are proposed, which enable the melding of non-parametric modeling results of multi-source data for the first time. Analytical melding results are derived to facilitate engineering applications. Finally, system reliability indices can be calculated, supporting system health management. The proposed methodology is applied to two real-world systems with different features of multi-source data, and detailed quantitative and comparative studies fully verify its effectiveness and superiority.