Adaptive categorization of complex system fault patterns

Due to large amount of information and the inherent intricacy, diagnosis in complex systems is a difficult task. This can be somehow simplified by taking a per-step towards categorizing the system conditions and faults. In this paper, the development and implementation of an approach that establishes class membership conditions, using a labelled training set, is described. More specifically, the use of negative recognition for classification and diagnosis of complex system faults are discussed. The adaptive recognition to achieve the classification is based on discovery of pattern features that make them distinct from objects belonging to different classes. Most of the existing approaches to fault diagnosis, particularly for large or complex systems, depend on heuristic rules. The approach proposed in this work does not resort to any heuristic rules, which makes it more suitable for diagnosis of faults in dynamic and complex systems. For evaluation purposes, using the data provided by the protection simulator of a large power system, its fault diagnosis is carried out. The results of those simulations are also reported. They clearly reveal that even for complex systems, the proposed approach, based on making use of the distinctive features of encountered fault patterns, is capable of fault classification with minimal supervision.