This dissertation research proposes embedding artificial neural networks into wireless sensor networks in parallel and distributed processing framework to implant intelligence for in-network processing, wireless protocol or application support, and infusion of adaptation capabilities. The goal is to develop in-network "intelligent computation" and "adaptation" capability for wireless sensor networks to improve their functionality, utility and survival aspects. The characteristics of wireless sensor networks bring many challenges, such as the ultra large number of sensor nodes, complex dynamics of network operation, changing topology structure, and the most importantly, the limited resources including power, computation, storage, and communication capability. All these require the applications and protocols running on wireless sensor network to be not only energy-efficient, scalable and robust, but also "adapt" to changing environment or context, and application scope and focus among others, and demonstrate intelligent behavior. The expectation from the research endeavor is to introduce computational intelligence capability for the wireless sensor networks to become adaptive to changes within a variety of operational contexts and to exhibit intelligent behavior.
The proposed novel approach entails embedding a wireless sensor network with an artificial neural network algorithm while preserving the parallelism and distributed nature of computations associated with the neural network algorithm. The procedure of embedding an artificial neural network, which may be configured for a problem either at wireless protocol or application levels, into the wireless sensor network hardware platform, which is a parallel and distributed processing system that is composed of a network of motes, is defined. This procedure is demonstrated for a case study with a Hopfield neural network and a minimum weakly connected dominating set problem as the model of wireless sensor network backbone or infrastructure. Issues and challenges pertaining to scalability, solution quality, and computational complexity for time and message are addressed through a comprehensive simulation study. Simulation study is performed using the TOSSIM environment for wireless sensor networks with mote counts up to 1000.
A comparative performance evaluation is performed. Solution quality, time and message complexity results for other centralized and distributed algorithms for connected dominating set construction as reported in the literature are used. Additionally, in-house simulation of non-distributed version of the proposed model is implemented to serve as a comparison benchmark and link to the studies in the literature. It is determined through the simulation study that the most critical factors that affect both the time complexity and the message complexity are the network size and time interval. The normalized computation time increases somewhat linearly for the most part for increases in the mote count the exception of the time interval value of 0.1 sec. The message complexity also increases with the increase in the mote count. The message complexity is not sensitive to the radio range but very sensitive to the time interval. All other parameters kept constant, the message complexity decreases with the increase in the time interval value on a consistent basis for all mote counts simulated. For smaller values of time interval, the network is more active due to motes waking up and exchanging messages more frequently, which leads to increased message complexity. The solution quality as measured by the size of the weakly connected dominating set by the proposed model is competitive with the performance exhibited by other algorithms reported in literature given all the adverse effects of computation being realized on a wireless sensor network platform. In light of the fact that there is significant opportunity to improve the entire wireless protocol stack for drastically reducing the time and space complexities through more efficient MAC, time synchronization and routing protocols, there is a strong prospect for the proposed architecture to scale up to tens of thousands of motes.