Application of artificial neural networks to synchronous generator condition monitoring
This thesis presents in Artificial Neural Networks(ANNs) based, automatic Pattern Recognition(PR) technique for electrical machines Condition Monitoring(CM) applications. The performance of synchronous generators has been studied under a variety of conditions and monitored using a range automatic pattern recognition approaches. The harmonic components of the generator stator, rotor and excitation currents have been analysed initially to gain information of fault conditions in the machines, and then as a source of data for input training patterns to the neural nets. Artificial neural networks; their architecture, algorithms and their application to pattern recognition have been studied. Two unsupervised self-organising neural networks were chosen for further investigation and applied to the automatic pattern recognition tasks. These two neural network models can be classified as Kohonen neural nets and Adaptive Resonance Theory nets. A computer implementation of Kohonen Self Organising Feature Maps(KSOFM) and a simulation that interprets the continuous valued model of adaptive resonance theory (ART2 net), have been studied in detail. General condition monitoring techniques for electrical machines have been briefly reviewed and statistical pattern recognition methods have also been described. To confirm the utility of the proposed ANNs based automatic pattern recognition techniques for electrical machine condition monitoring, two synchronous generators with different capacities, one of 8kva was used for training the networks, and another of 11kva for testing the networks, were employed in the experimental study. The stator, rotor and excitation current signals of a generator have been used to provide the networks' input patterns, and Kohenon networks and adaptive resonance theory networks applicability to electrical machines condition monitoring compared. The possibility of using the proposed techniques to real industrial systems has been discussed. Finally, some of the difficulties of implementing ANNs for condition monitoring are considered.