Excitation-based in-network processing for Wireless Sensor Networks

Motivation

Wireless sensor networks (WSNs) provide a promising and favorable solution for a variety of applications that require a dense or comprehensive long-term data collection. Due to the resulting amount of sensors, those networks often collect a significant amount of data. As communication is highly energy consuming and to prevent the network from congestion that might lead to inacceptable delays or packet loss it is required to apply some in-network processing data aggregation approaches.

Many applications in WSNs only need to detect an abnormal behavior and then notify other instances. These networks do not require detailed knowledge of each collected sensor value but instead a fast and reliable notification in case of unexpected data.

In an assisted living scenario, for example, it might be neither required nor intended to gather a detailed movement profile of each person. But it is important to be notified if the persons behavior changes unexpectedly – for example if he suddenly stumbles or falls while walking around or simply when getting out of the bed. Additionally, it might be useful to be able to identify groups of similar behavior, e.g. to distinguish those who are lying in bed from those who are walking around outside. With this general approach not only persons but also objects can be monitored such as the temperature curve of a fridge. Thus, an in-network data processing model is needed that both reduces communication and provides reliable and fast data routing to one or several sinks in case of unexpected sensor readings by any node inside the network.

For this purpose each node needs to be able to compute an abstract representation of its current environmental conditions by evaluating incoming sensor data. This abstract model should allow to validate further sensor readings and report if the deviation between expected and measured sensor data exceeds a certain threshold. The degree of deviation will be reported as an excitation level and can be used to determine routing priorities inside the network.

If the nodes continuously adjust their environmental model the amount of reported excitation and thus the network traffic is limited automatically. Additionally, flooding should also be prevented by the routing model. This can be reached by decreasing the routing priority of nodes that report a high excitation level over a longer period.

Based on the environmental models the nodes might also be able to build clusters of nodes sharing a similar environmental model. These clusters could be used for advanced routing strategies, more sophisticated data aggregation, or other kind of network intelligence.

Tasks

First of all a conceptual solution that covers the basic requirements of a sensor network as described above needs to be worked out. Therefore, existing approaches related to the different subtasks such as in-network data aggregation, environmental model generation, or routing should be taken into account. They can help to identify opportunities, challenges and pitfalls of the planned approach and guide further research and refinement. The different required conceptual and hierarchical building blocks need to be worked out and described to allow a targeted evaluation and potential realization.

In a following step a simple environmental representation for relatively static conditions should be implemented, e.g. based on mean value and standard derivation to monitor constant temperatures. As a target platform for all implementations Contiki OS should be considered to allow evaluation on several sensor platforms such as the INGA sensor node and on the network simulator Cooja. Knowledge gained from this basic sensing setup should be used to design an excitation level generation. It can also help to develop a flexible modular setup and guide further development of more sophisticated models.

To allow sending excitation data through a simple multi-hop network also a basic communication layer should be added. A sink may visualize incoming data for illustration, evaluation, and debugging purpose. Based on this basic setup more sophisticated approaches or algorithms can be designed. Depending on results from the previous examinations and the effort required, these algorithms can also be implemented and evaluated in a simulation or on a real sensor node.

In a more sophisticated routing the excitation level could be taken into account to examine priority-based routing approaches. One further aspect is to develop advanced environmental models that can deal with dynamic sensor readings such as data gained from an accelerometer mounted on a moving person. Here it would be important to find algorithmic solutions that are processable even on constrained sensor nodes.

Another aspect that deserves further investigation is the exchange of model data between sensor nodes and the comparability of different models. This can be used to enable nodes to verify data from other nodes or for behavior-based clustering approaches and advanced routing protocols as noted above.

If further promising fields of application or methods related to the topic are found during the work they should also be noted and can be analyzed deeper.