Automatic fault identification in sensor networks based on probabilistic modeling

Abstract

This work proposes a mechanism able to automatically categorize different types of faults occurring in critical infrastructures and especially water distribution networks. The mechanism models the relationship exhibited among the sensor datastreams based on the assumption that its pattern alters depending on the fault type. The first phase includes linear time invariant modeling which outputs a parameters vector. At the second phase the evolution of the parameter vectors is captured via hidden Markov modeling. The methodology is applied on data coming from the water distribution network of the city of Barcelona. The corpus contains a vast amount of data representative of nine network states. The nominal is included for enabling fault detection. The achieved classification rates are quite encouraging and the system is practical.