The myth and reality of wireless sensor networks: Designing optimally redundant networks

This paper considers some of the systems level issues concerned with the building of large scale sensing systems composed of autonomous intelligent sensors. The work is motivated by the potential offered by monolithic, integrated, intelligent MEMS sensors. They provide very low cost sensing in a form which may be included in 'smart structures' for example, in which structural diagnostic capability is deeply integrated into the structure itself. Designing and implementing such large sensing systems, however, raises a number of challenges, of which, many are still far from being resolved. Optimal redundancy, the follow on issue of fault management and their strategic consequences on the overall systems design requirements are chosen for discussion here. When considering the above, extended network lifetime is an important goal in the system design process.

A 'peer to peer' architecture of autonomous sensors, with no specialised control or processing nodes in the network is taken as the starting point of the discussion. This architecture is often put forward as being suitable for high reliability requirements, as the operation of the systems is not dependent on the operation of a single, specialised resource. Given that reliability generally depends on redundancy within the array an aim is the optimization of the level of redundancy in that array. Optimal redundancy depends on two systems functions. The first is a means by which a sensor in such an array may self detect fault conditions, and the second is the provision of a reliable means of handling these faults. A method of fault detection based on Artificial Neural Networks (ANNs) is described, which allows a sensor to detect whether it is functioning correctly, based on its own signal and that provided by a single neighbouring sensor. Within a decentralised, peer to peer network the identification of a neighbour and the subsequent handling of fault conditions require careful consideration, since, without a centralised controller, such systems are prone to deadlock conditions if the interaction between peers is not designed properly. The paper presents a protocol design for decentralised fault handling, modelled in the pi-calculus, a formalism which would allow the investigation of the dynamic properties of a system at the specification stage, as well as proof of the existence of required safety and liveness properties. Based on autonomous diagnostics, such as those described, and rigorous systems design techniques dependable networks of autonomous intelligent sensors may be designed.