Detection and Isolation of Failures in Directed Networks of LTI Systems

We propose a methodology to detect and isolate link failures in a weighted and directed network of identical multi-input multioutput linear time-invariant (LTI) systems when only the output responses of a subset of nodes are available. Our method is based on the observation of jump discontinuities in the output derivatives, which can be explicitly related to the occurrence of link failures. The order of the derivative at which the jump is observed is given by r(d+1), where r is the relative degree of each system's transfer matrix, and d denotes the distance from the location of the failure to the observation point. We then propose detection and isolation strategies based on this relation. Furthermore, we propose an efficient algorithm for sensor placement to detect and isolate any possible link failure using a small number of sensors. Available results from the theory of submodular set functions provide us with performance guarantees that bound the size of the chosen sensor set within a logarithmic factor of the smallest feasible set of sensors. These results are illustrated through elaborative examples and supplemented by computer experiments.