Michael Zoumboulakis

Escalation: complex event detection in wireless sensor networks

We present a new approach for the detection of complex events in Wireless Sensor Networks. Complex events are sets of data points that correspond to interesting or unusual patterns in the underlying phenomenon that the network monitors. Our approach is inspired from time-series data mining techniques and transforms a stream of real-valued sensor readings into a symbolic representation. Complex event detection is then performed using distance metrics, allowing us to detect events that are difficult or even impossible to describe using traditional declarative SQL-like languages and thresholds. We have tested our approach with four distinct data sets and the experimental results were encouraging in all cases. We have implemented our approach for the TinyOS and Contiki Operating Systems, for the Sky mote platform.

Active rules for sensor databases

Recent years have witnessed a rapidly growing interest in query processing in sensor and actuator networks. This is mainly due to the increased awareness of query processing as the most appropriate computational paradigm for a wide range of sensor network applications, such as environmental monitoring. In this paper we propose a second database technology, namely active rules, that provides a natural computational paradigm for sensor network applications which require reactive behavior, such as security management and rapid forest fire response. Like query processing, efficient and effective active rule execution mechanisms have to address several technical challenges including language design, data aggregation, data verification, robustness under topology changes, routing, power management and many more. Nonetheless, active rules change the context and the requirements of these issues and hence a new set of solutions is appropriate. To this end, we outline the implications of active rules for sensor networks and contrast these against query processing. We then proceed to discuss work in progress carried out in project Asene that aims to effectively address these issues. Finally, we introduce our architecture for a decentralized event broker based on the publish/subscribe paradigm and our early design of an ECA language for sensor networks.

Active rules for wireless networks of sensors & actuators

The last few years have witnessed a flurry of research in the field of query processing for networks of sensors and actuators. It is widely accepted that query processing is the method of choice for acquiring data from a sensor field. Although query processing offers a very good computational model for a variety of applications such as environmental monitoring, it is a poor match for application scenarios where a timely response to an event is required by the system. With this in mind, we propose a mature database technology, namely active rules, that provides a natural computational paradigm for sensor network applications that require reactive behavior, such as rapid forest fire response and security management. For the remainder of this paper we will outline the implications of active rules for sensor networks and contrast these against query processing. We will then proceed to discuss work in progress carried out by project <i>Asene</i> (Active SEnsor NEtworks) that aims to address these implications. We conclude by introducing our architecture for a decentralised event broker based on the publish/subscribe paradigm and our early design of an Event-Condition-Action (ECA) language for sensor networks.

Efficient Pattern Detection in Extremely Resource-Constrained Devices

We present a novel approach for the on-line detection of Complex Events in Wireless Sensor Networks. Complex Events are sets of data points that correspond to unusual patterns that can not be detected using threshold-based techniques. Our method uses an efficient implementation of SAX, a mature data mining algorithm, that transforms a stream of readings into a symbolic representation. Complex Event Detection is then performed via four alternative modes: (a.) multiple pattern detection using a suffix array, (b.) distance-based comparison, (c.) unknown pattern detection, and (d.) probabilistic detection. The method allows users to specify complex events as patterns or to search for interesting changes without supplying any information. The appropriateness of the approach has been verified by applying it to four sensor data sets. In addition, we have developed an efficient implementation for the TinyOS operating system, and further validated our assertions by collecting and analyzing data in real-time.

Estimation of Pollutant-Emitting Point-Sources Using Resource-Constrained Sensor Networks

We present an algorithm that makes an appropriate use of a Kalman filter combined with a geometric computation with respect to the localisation of a pollutant-emitting point source. Assuming resource-constrained inexpensive nodes and no specific placement distance to the source, our approach has been shown to perform well in estimating the coordinates and intensity of a source. Using local gossip to directionally propagate estimates, our algorithm initiates a real-time exchange of information that has as an ultimate goal to lead a packet from a node that initially sensed the event to a destination that is as close to the source as possible. The coordinates and intensity measurement of the destination comprise the final estimate. In this paper, we assert that this low-overhead coarse localisation method can rival more sophisticated and computationally-hungry solutions to the source estimation problem.

Integer-Based Optimisations for Resource-Constrained Sensor Platforms

In this paper we argue that the fundamental constrains of WSNs impose the need to re-discover programming optimisation techniques that were widely used a few decades ago but are less common today, at least in the conventional computing arena. Integer techniques, code tuning and profiling are absolutely essential in the world of the very small devices. We present three alternative methods of integer programming: scaling, fixed-point and rational arithmetic. These techniques are complemented by a brief review of bitwise and general optimisation techniques. As artifact of the usefulness of these techniques, we discuss the implementation details of a data mining algorithm that gained over a factor of 10 improvement in performance as a result of integer programming. We conclude by presenting a widely accepted time model adapted for a WSN platform.

Shared memories: a trail-based coordination server for robot teams

Robust, dependable and concise coordination between members of a robot team is a critical ingredient of any such collective activity. Depending on the availability and the characteristics of the particular communication infrastructure, coordination mechanisms can take varied forms, leading to distinct system behaviors. In this paper, we consider the case of robot teams operating within relatively sparse wireless sensor network deployments. We introduce Shared Memories, a trail-based coordination engine, that analyzes interaction patterns between participating team members and sensor network nodes capable to discover significant aggregate patterns, which are made available to the team. To this end, we propose a model for the representation of captured interactions and their sensory context developed as a probabilistic grammar, as well as associated metrics used to rank trails and quantify their significance. Such trails are used as the basis for coordinated operation in team tasks and are made available by the engine to all team members. Our implementation deploys ad-hoc wireless local networking capability available through surrogate devices to commodity robots and RFID proximity sensors. We report on the performance of this system in experiments conducted in a laboratory environment, which highlight the advantages and limitations of our approach.

Pattern Detection in Extremely Resource-Constrained Devices

Pervasive computing anticipates a future with billions of data producing devices of varying capabilities integrated into everyday objects or deployed in the physical world. In event-based systems, such devices are required to make timely autonomous decisions in response to occurrences, situations or states. Purely decentralised pattern detection in systems that lack time synchronisation, reliable communication links and continuous power remains an active and open research area. We review challenges and solutions for pattern detection in distributed networked sensing systems without a reliable core infrastructure. Specifically, we discuss localised pattern detection in resource-constrained devices that compriseWireless Sensor and Actuator Networks. We focus on online data mining, statistical and machine learning approaches that aim to augment decentralised pattern detection and illustrate the properties of this new computing paradigm that requires stability and robustness while accommodating severe resource limitations and frequent failures.

Ubiquitous Computing and Databases

One of the major challenges in turning the ubiquitous computing vision into reality is the development of distributed system architectures that will support effectively and efficiently the ability to instrument the physical world (Estrin et al., 2002; National Research Council, 2001). Such architectures are being developed around two core concepts: self-organizing networks of embedded devices with wireless communication capabilities and data-centricity. To augment physical artifacts with computational and communications capabilities, it is necessary to enable miniaturized hardware components capable of wireless communication. However, these same characteristics that allow for instrumentation of physical objects also impose significant constraints. Systems architectures require significant changes due to the severely limited resources available on these devices. One possible solution is offered by the emergence of data-centric systems. In this context, datacentric refers to in-network processing and storage, carried out in a decentralized manner (Estrin et al., 2000). One objective of data-centricity is to let systems exploit the anticipated high node densities to achieve longer unsupervised operating lifetimes. Indeed, smaller form factor wireless sensor nodes have limited resources and often cannot afford to transfer all the collected data to the network edge and forward to centralized information processing systems. A practical example of the data-centric approach for network routing is directed diffusion (Intanagonwiwat, Govindan & Estrin, 2000). This mechanism employs in-network processing by routing data along aggregation paths, thus, removing the need for an address-centric architecture. It exploits data naming as the lowest level of system organization and supports flexible and efficient in-network processing. In summary, databases have a dual role to play in ubiquitous computing: in the short term, they need to provide the mapping between physical and virtual entities and space in a highly distributed and heterogeneous environment. In the longer term, database management systems need to provide the infrastructure for the development of data-centric systems. Each of the two phases is discussed in turn in the following sections.