An efficient method for online detecting abnormal cascading pattern in distributed networks

In many large-scale real-time monitoring applications, such as water quality monitoring of large water distribution networks, massive streams flow out of multiple concurrent sensors continuously. Online detection of abnormal event, especially of those spreading in the area, is vital to such networks, as the event will influence a large number of nodes once breaking out. In this paper, we first define such event as abnormal cascading pattern, and propose an efficient, online approach to detection. Instead of analyzing the streams independently, we focus on the correlation among streams and its variation. We first summarize the evolving correlation between each pair of streams into a profile, distinguish the abnormal variation based on the profile, and then catch the cascading pattern through associating the abnormal pairs. Experiments indicate high detection sensitivity, low false alarm rate and background noise tolerance of our approach.