Performance of data aggregation for wireless sensor networks

This thesis focuses on three fundamental issues that concern data aggregation protocols for periodic data collection in sensor networks: which sensor nodes should report their data, when should they report it, and should they use unicast or broadcast based protocols for this purpose. The issue of when nodes should report their data is considered in the context of real-time monitoring applications. Such applications can require high sampling rates and low-delay forwarding of the sensor values to a sink node at which the data is to be further processed. Since aggregation requires that some sensor data be delayed at intermediate nodes, however, while waiting for other data to be received, a key issue in the context of real-time monitoring is how to achieve effective aggregation with minimal forwarding delay. Previous work has advocated synchronous aggregation, in which a node’s position in the aggregation tree determines when it transmits to its parent. The first part of this thesis shows that asynchronous aggregation, in which the time of each node’s transmission is determined adaptively based on its local history of past packet receptions from its children, outperforms synchronous aggregation by providing lower delay for a given end-to-end loss rate. Second, new broadcast-based aggregation protocols are designed and evaluated. They minimize the number of packet transmissions, relying on multipath delivery rather than automatic repeat request for reliability. The performance of broadcast-based aggregation is compared to that of unicast-based aggregation, in the context of both real-time and delay-tolerant data collection. For real-time applications, this work investigates whether such protocols can achieve lower collection delays and support higher sampling rates than conventional aggregation protocols, while performance evaluation for delay-tolerant data collection is focused on reliability. The results suggest that when packet loss is random, broadcast-based protocols can yield significantly improved performance in some real-time data collection scenarios, specifically when sensor data can be aggregated into packets of size that is independent (or largely independent) of the number of values being aggregated. Broadcast-based aggregation can yield significantly better performance than unicast-based aggregation for both real-time and delay-tolerant data collection when packet loss follows a two-state Gilbert error model. Finally, in the context of applications in which coverage of some monitored region is to be maintained, this thesis investigates the potential benefits of dynamically, rather than semi-statically, determining the set of nodes reporting their data. In such applications, sensor nodes are often deployed more densely than would minimally be required. With a semi-static approach, node scheduling protocols are deployed to reduce energy consumption and prolong network lifetime by putting redundant nodes to sleep. Node scheduling approaches, however, may leave part of the monitored region uncovered if node failures happen and a replacement node is not woken up immediately. Unicast and broadcast-based coverage-preserving data aggregation protocols in which nodes dynamically determine during each round of data collection whether they should transmit their data, or whether the set of neighbouring nodes that have already transmitted is sufficient to provide coverage, are designed and evaluated. The performance of the proposed protocols is compared to that of data collection protocols relying on node scheduling. Results suggest that the proposed broadcast-based protocol can greatly improve reliability, at the potential cost of increased traffic volume owing to non-minimal selection of transmitting nodes. For real-time data collection that is willing to trade reliability for improved data collection delay, broadcast-based data aggregation with node scheduling is able to achieve lower delay with moderate loss of coverage.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.