Abstract
Wireless sensor networks (WSNs) are basically unreliable data gathering systems that utilizes a large number of small, battery-powered, and resource-limited sensing devices to monitor certain/various phenomena in the real world. These systems can contribute to intelligent services such as, surveillance, healthcare, environmental and utility monitoring. In WSNs, hundreds or thousands of these disposable sensor nodes are connected wirelessly to communicate either among each other or directly to the base station. How to reduce the power consumption and also lengthen the system life time is one of the key issues to sustain the services effectively. According to the radio model, packet transmission is depletes a much more substantial amount of the energy budget when compared to sensing and processing. Therefore, it is desirable to compress or filter the sensing data to a minimum in order to reduce the transmission power required. The model-based scheme is a promising solution of data compression for WSNs because the data streams are highly correlated temporally. In this thesis, a tree-structured linear approximation with optimal rate-distortion (RD) control method to deal with the temporal data stream is proposed. This method approximates temporal data by a piecewise linear function, consisting of connected line segments. Initially, a set of possible line segments of equal size are maintained as a complete binary tree for N consecutive data of current stream in each of sensor nodes. Next, a RD pruning process is applied to trim the tree to several candidate forms of incomparable RD pairs. Then, an optimal distortion allocation procedure is employed to allocate the distortions to sensor nodes accordingly. With the assigned distortion value, each tree is further shrunken by iteratively merging possible pair of segments to a minimum rate (# of line segments) while obeying the distortion. Finally, a refinement procedure, with the assistance of outlier removal, is implemented to further compress the data streams. A real life data set is applied to demonstrate the effectiveness of the thesis. For nearly all combinations with distortion requirements, the proposed method outperforms the earlier approaches in terms of data reduction.