Title:Quickest detection of cascading failure

Abstract:We consider online detection of cascading failure in the networks using sequential data. Our goal is to detect the failure as quickly as possible after it occurs. To achieve this goal, we propose a temporal diffusion network model to capture the dynamic of the potential change, which will help us to capture the change as quickly as possible since its onset. Under this model, the hazard rate of a node increases as the number of failure neighbors increases. Once the failure affects a node, the distribution of the measurement changes from pre-change distribution to an unknown post-change distribution. We develop a sequential generalized likelihood ratio statistics, which performs joint detection and estimation in detecting the change. Numerical experiments show that our method outperforms the existed methods.