A Clustering Algorithm to Improve the Scan Statistic in Sensor Detection Systems

The success of the scan statistic in detecting anomalies in georeferenced data has motivated its use in distributed sensor systems to detect an emitter at an unknown location. Sensors are grouped into clusters, cluster statistics are produced, and the scan statistic decides that the emitter is present if any cluster statistic is above a threshold. Although the scan statistic is not the optimal fusion rule, it avoids combining strong measurements from sensors near the emitter with weak measurements from sensors far from the emitter. The question that motivates this paper is: could a clustering algorithm improve the detection performance of the scan statistic? Previous studies on the scan statistics considered that the set of clusters is given or is the product of a scanning process; and previous studies on clustering algorithms for wireless sensor networks have not considered forming clusters specifically for the scan statistic. Our first goal is to highlight the opportunity of improving the scan statistic by carefully designing the cluster set. We discuss how the cluster set influences not only the detection performance, but also processing and communication in the system. Our second goal is to propose and study a new clustering algorithm to build the cluster set for the scan statistic. Although suboptimal, our algorithm produces cluster sets that reach similar or better detection performance than the usually considered cluster sets with a significantly lower number of clusters, which results in less processing and communication in the system.