Date of Award
Doctor of Philosophy (PhD)
Michele C. Weigle
Irwin B. Levinstein
Dimitrie C. Popescu
Sensor networks are deployed to monitor a seemingly endless list of events in a multitude of application domains. Through data collection and aggregation enhanced with data mining and machine learning techniques, many static and dynamic patterns can be found by sensor networks. The aggregation problem is complicated by the fact that the perceived value of the data collected by the sensors is affected by many factors such as time, location and user valuation. In addition, the value of information deteriorates often dramatically over time. Through our research, we already achieved some results:
A formal algebraic analysis of information discounting, especially affected by time. A general model and two specific models are developed for information discounting. The two specific models formalize exponetial time-discount and linear time-discount. An algebraic analysis of aggregation of values that decay with time exponentially. Three types of aggregators that offset discounting effects are formalized and analyzed. A natural synthesis of these three aggregators is discovered and modeled. We apply our theoretical models to emergency response with thresholding and confirm with extensive simulation. For long-term monitoring tasks, we laid out a theoretical foundation for discovering an emergency through generations of sensors, analysed the achievability of a long-term task and found an optimum way to distribute sensors in a monitored area to maximize the achievability. We proposed an implementation for our alert system with state-of-art wireless microcontrollers, sensors, real-time operating systems and embedded internet protocols.
By allowing aggregation of time-discounted information to proceed in an arbitrary, not necessarily pairwise manner, our results are also applicable to other similar homeland security and military application domains where there is a strong need to model not only timely aggregation of data collected by individual sensors, but also the dynamics of this aggregation. Our research can be applied to many real-world scenarios. A typical scenario is monitoring wildfire in the forest: A batch of first-generation sensors are deployed by UAVs to monitor a forest for possible wildfire. They monitor various weather quantities and recognize the area with the highest possibility of producing a fire --- the so-called area of interest (AoI). Since the environment changes dynamically, so after a certain time, the sensors re-identify the AoI. The value of the knowledge they learned about the previous AoI decays with time quickly, our methods of aggregation of time-discounted information can be applied to get update knowledge. Close to depletion of their energy of the current generation of sensors, a new generation of sensors are deployed and inherit the knowledge from the current generation. Through this way, monitoring long-term tasks becomes feasible. At the end of this thesis, we propose some extensions and directions from our current research:
- Generalize and extend the special classes of Type 1 and Type 2 aggregation operators;
- Analyze aggregation operator of Type 3 and Type 4, find some special applicable candidates;
- Data aggregation across consecutive generations of sensors in order to learn about events with discounting that take a long time to manifest themselves;
- Network implications of various aggregation strategies;
- Algorithms for implementation of some special classes of aggregators.
- Implement wireless sensor network that can autonomously learn and recognize patterns of emergencies, predict incidents and trigger alarms through machine learning.
"Towards Aggregating Time-Discounted Information in Sensor Networks"
(2016). Doctor of Philosophy (PhD), Dissertation, Computer Science, Old Dominion University, DOI: 10.25777/h0vm-d468