A data-driven approach for predicting failure scenarios in nuclear systems

A data-driven approach is presented for the on-line identification of the system Failure Mode (FM) and the prediction of the available Recovery Time (RT) during a failure scenario, i.e., the time remaining until the system can no longer perform its function in an irreversible manner. The FM identification and RT prediction modules are linked in a general framework that recognizes the patterns of dynamic evolution of the process variables in the different system failure modes. When a new failure scenario develops, its evolution pattern is compared by fuzzy similarity analysis to a library of reference multidimensional trajectory patterns of process variables evolution; the failure mode of the developing scenario is identified by combining the modes of failure of the reference patterns, weighed by their similarity to the developing pattern; the similarity weights are then fed to the RT prediction module that estimates the time remaining before the developing trajectory pattern hits a failure threshold.