Heterogeneous sensor fusion in sensor networks: A language-theoretic approach

This dissertation presents a framework for feature-level heterogeneous sensor data fusion in sensor networks via a language-theoretic approach. Probabilistic finite state automata (PFSA) are used to model the semantic patterns in the observations of the sensors. A novel pattern discovery algorithm is developed to extract the PFSA model from symbol sequences. It is shown that this algorithm can capture semantic structures more effectively than the existing techniques. In order to formulate the data fusion problem for semantic features, a link is established between the formal language theory and functional analysis by constructing a Hilbert space over a class of stochastic regular languages represented by PFSA. New algebraic operations are defined for PFSA with a family of parametrized inner products. The norm induced by the inner product is interpreted as a measure of the information contained in PFSA. Applications of this technique are discussed in the following areas: a) Orthogonal projection in the Hilbert space to solve the model reduction problem of PFSA. Numerical examples and experimental results are provided to elucidate the process of model order reduction. b) Supervised learning of semantic features of heterogeneous sensor data in the product Hilbert space. The semantic features are combined optimally for classification using linear discriminant analysis (LDA). The proposed algorithm has a set of parameters that can be potentially configured by the users to adapt the algorithm to environment changes. The proposed algorithm is validated for object recognition at the US-Mexican border. An architecture of fusion-driven sensor networks is introduced to incorporate the proposed fusion framework in sensor networks. The network protocol, called dynamic time-space clustering (DSTC), and its heterogeneous version, are designed to adapt the network to the fusion algorithms. A sensor network for selectively tracking mobile targets is implemented in the network simulator NS-2 for both homogeneous and heterogeneous sensor fields in an urban scenario for validating the propose architecture.