Distributively Observed Events in Wireless Sensor Networks
Norman Dziengel · Universitätsbibliothek der FU Berlin Hochschulschriftenstelle u. Dokumentenserver · 2016
Wireless Sensor Network (WSN)s comprise multiple battery powered and wireless networking and sensor based minicomputers called Sensor Node (SN)s. SNs are typically equipped with at least one sensor to autonomously observe the environment by acquiring sensor data that is processed in the embedded hardware of the SNs in order to potentially communicate data to other wirelessly connected SNs of the network. Event Detection is an observing and assessing process of real incidents or phenomena. WSNs have the potential to observe and detect environmental events in order to offer support in safety matters such as Structural Health Monitoring (SHM) to detect e.g. age-related bridge damage, areal overviews to support firefighting operations, or fence monitoring systems to detect intruders. Event detection with WSNs is challenging because events typically cause different measurements at SNs at different locations, but we want a single comprehensive meaning or interpretation as a result. A simple strategy based on redundant data collection makes sense in case of threshold based events such as fire detection, whereas pattern based events like intrusion detection at fences benefit from multiple perspectives on the event. The requirements for event- observing WSNs are in contradiction to their properties, especially if they need to sustain their functionality over a long period and need to deliver accurate event detection while using small and ubiquitous sensor devices with limited energy and limited computational power. This thesis presents a Distributed Event Detection system that shifts the evaluation process of the event data from the Base Station (BS) into the network. This distributed evaluation concept reduces the communication load to one single event notification which leads to an increased lifetime of the most charged SNs as well as the whole network. We introduce two frameworks that help to realize the concept of the Distributed Event Detection, while having the goal to preserve the global event knowledge. The frameworks make this possible by preserving the diverse event perspectives of the SNs despite a necessary data reduction. The Evaluation Framework comprises the automated creation of an event model based on the data of a supervised training to support the transferability of the system on most different applications. The Distributed Event Detection Framework enables real world deployments by implementing the detection system on the SNs that uses the trained model within the wireless network. The SNs affected by the event autonomously exchange event information within the network to cooperatively classify the event. In the end the final decision is made on whether the detected event is worth notifying the BS or any other responsible system for. We present the technical limits for our system by investigating the trade-off between the communication savings to the BS and the increased in-network communication. Compared with other classical information fusion approaches, we can clearly attest an outperforming energy efficiency due to the conceptionally reduced communication. All compared classical information fusion approaches are conceptually reused and modularly extended within the Distributed Event Detection system in order to ensure that our proposed system runs under a non-recommended setup as it can fall back step-wise to the underlying information fusion concepts to support the needs of the application. The proposed Distributed Event Detection System reaches a high event detection sensitivity of more than 80% up to 100% depending on the application. With four real world deployments, we show the functionality, event detection performance, and the application specific lessons learned, which are evaluated in relation to the Distributed Event Detection’s applicability.