Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


Computer Science

First Advisor

S. S. Iyengar


A system reacting to its environment requires sensor input to model the environment. Unfortunately, sensors are electromechanical devices subject to physical limitations. It is challenging for a system to robustly evaluate sensor data which is of questionable accuracy and dependability. Sensor fusion addresses this problem by taking inputs from several sensors and merging the individual sensor readings into a single logical reading. The use of heterogeneous physical sensors allows a logical sensor to be less sensitive to the limitations of any single sensor technology, and the use of multiple identical sensors allows the system to tolerate failures of some of its component physical sensors. These are examples of fault masking, or N-modular redundancy. This research resolves two problems of fault masking systems: the automatic calibration of systems which return partially redundant image data is problematic, and the cost incurred by installing redundant system components can be prohibitive. Both are presented in mathematical terms as optimization problems. To combine inputs from multiple independent sensors, readings must be registered to a common coordinate system. This problem is complex when functions equating the readings are not known a priori. It is even more difficult in the case of sensor readings, where data contains noise and may have a sizable periodic component. A practical method must find a near optimal answer in the presence of large amounts of noise. The first part of this research derives a computational scheme capable of registering partially overlapping noisy sensor readings. Another problem with redundant systems is the cost incurred by redundancy. The trade-off between reliability and system cost is most evident in fault-tolerant systems. Given several component types with known dependability statistics, it is possible to determine the combinations of components which fulfill dependability constraints by modeling the system using Markov chains. When unit costs are known, it is desirable to use low cost combinations of components to fulfill the reliability constraints. The second part of this dissertation develops a methodology for designing sensor systems, with redundant components, which satisfy dependability constraints at near minimal cost. Open problems are also listed.