Data fusion on a distributed heterogeneous sensor network

Alarm-based sensor systems are being explored as a tool to expand perimeter security for facilities and force protection. However, the collection of increased sensor data has resulted in an insufficient solution that includes faulty data points. Data analysis is needed to reduce nuisance and false alarms, which will improve officials' decision making and confidence levels in the system's alarms. Moreover, operational costs can be allayed and losses mitigated if authorities are alerted only when a real threat is detected. In the current system, heuristics such as persistence of alarm and type of sensor that detected an event are used to guide officials' responses. We hypothesize that fusing data from heterogeneous sensors in the sensor field can provide more complete situational awareness than looking at individual sensor data. We propose a two stage approach to reduce false alarms. First, we use self organizing maps to cluster sensors based on global positioning coordinates and then train classifiers on the within cluster data to obtain a local view of the event. Next, we train a classifier on the local results to compute a global solution. We investigate the use of machine learning techniques, such as k-nearest neighbor, neural networks, and support vector machines to improve alarm accuracy. On simulated sensor data, the proposed approach identifies false alarms with greater accuracy than a weighted voting algorithm.