National Competence Center in Research for Mobile Information and Communication Systems (NCCR-MICS)

Abstract

The NCCR-MICS project was launched in 2001. Its goal is to study fundamental and applied research questions raised by new generation mobile communication and information services, based on self-organization. Such systems have become very topical with the advent of mobile ad-hoc, peer-to-peer, and sensor networks. The 2nd phase of NCCR-MICS is composed of more than twenty research projects distributed over four clusters. The research project of the RVS group on "Distributed event detection and localization architecture for wireless sensor networks" (IP4) aims at designing and implementing a distributed event detection, event localization, and event classification framework. It includes efficient and reliable signaling protocols as well as mechanisms to dynamically reconfigure its specific sensor network applications.

We investigated the classification of discrete events, computed on tiny wireless sensor nodes. Three different classifiers have been evaluated: a Bayesian classifier, a fuzzy logic controller, and a neural network approach. We assume that no a priori knowledge about the event classes is available and events are only observable as collections of raw sensor data. Accordingly, event classes need to be learned from that raw training data. In our work, event classes are learned by a k-means clustering algorithm. Any subsequent classifier training is based on these extracted event classes. Thus, the resulting classifiers are completely self-learning. Event classes are learned from emitted signal strength estimations, which are collected and processed by dynamically established tracking groups. The resulting event estimates are reported to a base station, where the classifiers are trained. The learned classifier parameters are then downloaded onto the sensor nodes, where any subsequent classification and filtering is performed.

Furthermore, we developed a node-level decision unit of a self-learning anomaly detection mechanism for office monitoring with wireless sensor nodes. The node-level decision unit is based on Adaptive Resonance Theory (ART), which is a simple kind of neural networks. The Fuzzy ART neural network used is an adaptive memory that can store a predefined number of prototypes. Any observed input is compared and classified in respect to a maximum number of M online learned prototypes. The Fuzzy ART neural network is used to process, classify, and compress time series of event observations on sensor node level. Based on simple computations, each node is able to report locally suspicious behaviour. A systemwide decision is subsequently performed at a base station. The system has been used for the detection and reporting of abnormal building access with a wireless sensor network. An office room, offering space for two working persons, has been monitored with ten sensor nodes and a base station. The task of the system is to report suspicious office occupation such as office searching by thieves. On the other hand, normal office occupation should not throw alarms. In order to save energy for communication, the system provides all nodes with some adaptive short-term memory. Thus, a set of sensor activation patterns can be temporarily learned. Unknown event patterns detected on sensor node level are reported to the base station, where the system-wide anomaly detection is performed. The anomaly detector is lightweight and completely self-learning. The system can be run autonomously or it could be used as a triggering system to turn on an additional high-resolution system on demand. Our building monitoring system has proven to work reliably in different evaluated scenarios. Communication costs of up to 90% could be saved compared to a threshold-based approach without local memory.

Moreover, research on sensor MAC and routing protocols for sensor-based distributed monitoring applications has been performed. Contention-based MAC protocols following periodic listen/sleep cycles face the problem of virtual clustering if different unsynchronized listen/ sleep schedules occur in the network. Border nodes, which maintain all respective listen/sleep schedules, are required to interconnect these virtual clusters. This is however a waste of energy, if locally a common schedule can be determined. We propose to achieve local synchronization with a mechanism that is similar to gravitation. Clusters represent the material, whereas synchronization messages sent by each cluster represent the gravitation force of the according cluster. Due to the mutual attraction caused by the clusters, all clusters merge finally. Moreover, we developed a routing backbone construction mechanism that exploits and uses the synchronization messages exchanged by synchronized contention-based MAC protocols. Due to the usage of synchronization messages no additional control traffic is required to setup the routing backbone. Every node running a synchronized contention-based MAC protocol follows a given listen/sleep cycle. Because routing is supported by the backbone, non-backbone nodes can temporarily turn off their radios for multiple listen/ sleep cycles. Thus, additional energy can be saved. Accordingly, nonbackbone nodes do not have to wake up in every listen/sleep cycle to synchronize with other nodes, but wake up only if they have to report some sensor readings to a base station. In this case, they synchronize to the backbone, send their data, and go back to sleep after successful transmission. Our approach is applicable to rather static networks with mainly source-to-sink traffic. Most monitoring applications are of this kind.

Project Details