Abstract
Currently, most of the sensors are ``smart'' in nature, which means that sensing elements and associated electronics are integrated on the same chip. Among these new generation of sensors, the Micro-Electro-Mechanical-Systems (MEMS) make use of Microelectronics technologies for batch manufacturing of small footprint sensors to unprecedented volumes and at low prices. If those components of the shelf are satisfactory for many consumer and low- to medium-end applications, they still cannot fully meet the performance needs of many high-end applications.However, due to their decreasing price, their small footprint, and their low-power consumption, it is now feasible to implement systems with tens and even hundreds of sensors. Those systems give a possible solution to the lack of performance of individual sensors and additionally they can also improve dependability and robustness of sensing. Sensor array systems are one of these methods of redundant measurements that arise in response to the aforementioned problems. The development of data fusion algorithms for sensor array systems is a research topic frequently studied in the literature. Even so, it still remains a lot of research work to do in this increasingly important area. The emergence of new applications with increasingly complex needs is growing the requirement for new algorithms with features such as integration, adaptability, dependability, low computational cost, and genericity among others.In this thesis we present a new algorithm for sensor array systems that propose a viable solution to overcome constraints mentioned before. The proposal is an on-line method based on the MInimum Norm Quadratic Unbiased Estimation (MINQUE) that is able to compute sensors' variances without the knowledge of the inputs. This algorithm is capable to track changes in sensors' variances caused principally by the low-frequency noise effects, as well as to detect and point out sensors affected by permanent or transitory errors. This approach is generic, which means that it can be implemented for different types of sensor array systems. In addition, this algorithm can be also implemented in sensor network systems.Two more contributions of this thesis can be listed. The first is a generic sensor model for sensor simulations at system level. This tool created inside the Matlab Simulink environment permits the analysis of implementations of data fusion algorithms in multi-sensor systems. Unlike the models previously existing in the literature, this sensor model has characteristics such as genericity and inclusion of low-frequency noises. The second is a study to compare the performance and feasibility in the implementation of different algorithms for data fusion in sensor array systems. This study contains an analysis of computational complexity, memory required, and the error in estimation. The analyzed algorithms are : the method of least squares, an artificial neural network, Kalman filter, and Random weighting.