Descoberta de padrões de alarme redundantes com técnicas de mineração de dados e redes complexas

Abstract

Alarm management is a research area that is growing rapidly on industrial automation topics. One of the major challenges in alarm rationalization, in which the volume of generated alarms is reduced to an appropriate number so that a human being can handle them, is to identify, between files and databases containing tens of thousands of daily records, patterns that might indicate unnecessary alarms. This work presents a new approach which combines sequence mining, association rules extraction with MNR (Minimum Non Redundant Association Rules), cross-correlation analysis, and complex network modeling for visualization, creating a more comprehensive alternative to the detection process. The solutions performance in terms of accuracy shows improvements over the best existing approach, resulting in an more reliable and predictable alternative for the identification of meaningful patterns.