Fault Localization Upon Non-Supervised Neural Networks and Unknown Input Observers for Bounded Faults

The task of fault diagnosis consists of determining the type, size and location of the fault as well as its time of detection. The use of knowledge-based techniques for fault localization and diagnosis allows on-line recognition of abnormal scenarios. These are based upon data treatment (Nelles, 2001), albeit these techniques require large amounts of data in order to obtain a valid representation of different scenarios. Alternatively, analytical redundancy allows a highly accurate detection of faults, based on a model of the observed system. Nevertheless, analytical redundancy requires a very accurate model of the system in order to locate a fault. Both, knowledge-based techniques and analytical redundancy, allow localization and classification of unknown scenarios as abnormal situations. The advantages of both methods depend on the type of information obtained, such as heuristic knowledge or model-based implementation. However, for abnormal situations, they have the disadvantage of not providing accurate results. In general, both methods require two important features: (a) the capability to determine faults, and (b) its sources of information. Several different approaches attempt to combine knowledge-based techniques and analytical redundancy. (Venkatasubramanian V., et al., 2003a) (Venkatasubramanian V., et al., 2003b) (Venkatasubramanian V., et al., 2003c) present an extended overview of fault localization and diagnosis based on modeland knowledge-based techniques. In general, the combination of both methods is feasible, although presenting undesirable glitches when used simultaneously, as discussed by (Liling et al., 2002). Several approximations have reviewed this constraint like (Su T., et al., 2008) where function approximation is pursued using a hybrid artificial neural network where data analysis becomes crutial for this purpose. In a similar manner (Zhong & Wang, 2008) presents a support vector regression where data uncertainty is studied, giving a good idea about the inherent characteristics of the data in order to by analysed. Several issues need to be addressed in order to study data analysis for system identification and representation, most of there are out of the scope of this paper. The goal of the approach followed here is to