Stochastic Simulation Method for Reasoning of Dynamic Uncertain Causality Graph (DUCG)

Abstract

Dynamic Uncertain Causality Graph (DUCG) is an innovative model developed recently on the basis of dynamic causality diagram (DCD) model, which has been proved to be reliable for fault diagnosis of nuclear power plants. DUCG can represent complex uncertain causal relationship graphically, with both high efficient inference and support of incomplete expression. Therefore, DUCG is often built much larger than Bayesian Network (BN). However, as the scale of real problem is so large, DUCG still has the problem of combination explosion. Stochastic Simulation is a common solution for it. However, it is almost impossible to use traditional sampling algorithms for DUCG because the joint probability of evidences could be less than 10−20. In this paper, the algorithm based on conditional stochastic simulation for the inference of DUCG was proposed. It obtains the probability of evidences by calculating the expectation of the conditional probability in sampling process instead of using the sampling frequency, which overcomes the difficulty. What’s more, this algorithm uses recursive reasoning method of DUCG to calculate conditional probability distributions of node for sampling, which means this process only depends on its parent nodes’ states. As a result, the algorithm features in lower time complexity. In addition, it has the potential of parallelization like other sampling algorithms. In conclusion, this algorithm is promising to provide a new solution to the inference of the DUCG in large-scale and complex state situations.