Robust Real-Time Event Services in Wireless Sensor Networks

Event detection is one of the main components in numerous wireless sensor network (WSN) applications. Regardless of the specific application, the network should be able to detect if particular events of interest have occurred or are about to. Traditional event services allow for the definition of events including correlated events, registering for events, and upon occurrence of events, detection and notification of events. In WSNs, events are not binary, but are based on sensor fusion from many noisy sensors in complicated environments. Sensor data may be missing, wrong, or out of date. Consequently, event services must operate in real-time, support data fusion and confidence calculations, and conserve power. Event services must also fundamentally recognize location, since sensor network events are a function of where they occur. Sensor network event services must be highly decentralized in order to work on the limited capacity devices. The services must also minimize false alarms. All these features make building event services for WSNs very challenging.

This research focuses on enabling the design and development of robust real-time event services. More specifically this work investigates the following research problems:

Event specification: We have developed a formal event description language which is an enhanced Petri net and combines features from Colored, Timed and Stochastic Petri nets. This language, coMpact Event Detection and Analysis Language (MEDAL), can capture the structural, spatial, and temporal properties of a complex event detection system. MEDAL also addresses key aspects of sensor networks, such as communication, actuation, and feedback control. MEDAL's graphical support, inherited from Petri nets, makes the application models easy to understand and accessible to a wide range of users.

Event detection: The majority of current event detection approaches rely on using precise, also called "crisp", values to specify WSN events. However, we believe that crisp values cannot adequately handle the often imprecise sensor readings. In this work we have studied how using fuzzy values could improve the accuracy, timeliness, and resource requirements of event detection. Our experiments with real-world fire data have shown that using fuzzy values results in more accurate event detection than when crisp logic is used, since fuzzy logic is more resilient to imprecise sensor readings.

Robustness to node failures: Even if an event detection system has been correctly designed and built, its continuous and reliable operation is difficult to guarantee due to hardware degradation and environmental changes. However, not all node failures have the same effect on applications' behavior. Some node failures are critical and lead to significant application degradation, while others may not affect the application at all. We have designed techniques to detect node failures that affect the application-level behavior of the system and minimize the number of maintenance dispatches without sacrificing the event detection accuracy of the application.

The techniques and approaches presented in this dissertation have the potential to:

1. Broaden the range of events detected by sensor network applications.

2. Provide formal specification and broad analysis capabilities for event-driven sensor network applications.

3. Improve the accuracy of event detection.

4. Help verify at runtime that a WSN application is able to satisfy its high-level requirements.

5. Improve the accuracy of event detection applications in the presence of node failures.

6. Decrease the number of necessary maintenance dispatches needed to preserve satisfactory application behavior.