Sensor fault detection using fuzzy logic and neural networks

The validation of signals is a technique which integrates information from redundant and from functionally diverse sensors to provide highly reliable information to operating crews and to automatic controllers. Signal validation is generally performed by like-sensor comparisons (direct redundancy). When the increase of sensors is impossible, one uses in preference analytical redundancy. Analytical redundancy refers to the physical relationships, such as conservation of mass or conservation of energy, that exist among the many variables being measured in a system. These validation techniques must be implemented to increase systems reliability and to facilitate detection of failures. This paper presents several methods of detection and location of measurement failures using or not models of the process. For the model-based approach the authors present the standardized imbalance residuals analysis and the standardized least square residuals; usually, these techniques are presented by using linear models and here the authors extend to non-linear models. The authors then compare these approaches with fault detection using neural nets