Modeling systems in a Dynamic Bayesain Network is a powerful technique for uncertainty quantification, updating, data fusion, and forecasting which can be used in many practical applications. A modular and customizable network architecture along with a suite of statistical tools is flexible enough to accommodate many engineering problems from diverse fields such as design for reliability, risk assessment, maintenance and inspection scheduling, experiment design, and more. A common and flexible filtering technique for solving a DBN is sequential importance resampling (SIR). This sampling-based method continually adjusts the samples to lie in the most likely region of the parameter space. SIR can obtain the sampling efficiency of importance sampling while overcoming the need for user-specified importance distributions. The method can, however, result in noisy outputs which are functions of higher moments of the posterior distribution. For example, estimates of the tail probabilities of the updated model do not converge well. While sequential importance sampling (SIS) preserves the tail probabilities better, it requires expert user input to define the importance distribution. In this work, we develop two modifications to the resampling in SIR to improve convergence of quantities of interest that depend on higher moments of the posterior. The first is to perform weighted sampling multiple times until a good solution is reached. The second is a constrained resampling method which optimizes the length scale distribution in kernel density-based resampling to preserve the quantity of interest in the posterior samples. The improved method is demonstrated on an example problem of computing a posterior tail probability, namely the hazard rate.
展开▼
