(Data obrony: 2011) Kozik, Jarosław Wydział Elektrotechniki, Automatyki, Informatyki i Elektroniki
The thesis presents the results of diagnostic investigations of four different faults in a synchronous machine: a short circuit in a coils group of armature winding, a break in one armature parallel branch, as well as short-circuit and break in the pole winding of excitation circuit. For the purpose of research a test bench and specialized measuring equipment was constructed. In order to improve the competitiveness of the investigations a great attention was put on diagnosing the early stages of damage, thus the faults involving short-circuits were conducted with a gradually increasing intensity. The measurements were followed by a spectral analysis of the recorded waveforms and choosing the components present in the spectra, that can be considered as the symptoms of a particular fault. The investigations also included spectral analysis of the Park's vector and its 100Hz component. Then the fault symptoms were determined fully automatically using a technique called feature selection. The problem was presented as a searching in a multidimensional feature space such a subspace (represented by a set of components of signal spectrum), which would give the greatest difference between the healthy and the faulty states. As a criterion of this condition a Mahalanobis distance measure was used. Due to the relatively large size of the search space an efficient search algorithm was required. In this role the simple genetic algorithm (SGA) was used, which was then modified by adding a few improvements. Designated symptoms were used as input for two different types of classifiers: neural and fuzzy, whose parameters were tuned using a genetic algorithm.
