Comparison of genetic and incremental learning methods for neural network-based electrical machine fault detection

Learning of neural network structure and parameters using genetic and incremental heuristic algorithms are potential approaches to address the local optima and design issues experienced when using conventional deterministic algorithms and arbitrarily chosen network structures. This chapter presents results on the development of an evolutionary (EANN) and an evolving fuzzy granular (EGNN) neural network for detecting and classifying inter-turns short-circuit in the stator windings of induction motors. A condition monitoring system based on features extracted from voltage and current waveforms associated with a type of neural network is proposed. Both EANN and EGNN are able to develop their parameters and structure according to data samples and a fitness function or error index. Real-programming-based genetic operators, i.e., mutation, recombination, and selection operators, were customized to address the fault detection problem in question. Operators and rates were evaluated in order to obtain a consistent and effective EANN learning algorithm. EGNN handles gradual and abrupt changes typical of nonstationary environment using fuzzy granules and fuzzy aggregation operators. While EGNN is online adaptive, EANN requires a set of initial data. Such initial dataset for training offline neural networks was obtained from a modified induction motor properly designed for insertion of stator shorted-turns and from a fault simulation model useful to extend the dataset by interpolation. Aspects of the neural models, such as classification performance, computational complexity, and compactness, are compared with each other and with the results obtained using a conventional feedforward neural network of similar structure, but trained by a deterministic gradient descent algorithm. The EGNN classifier achieved the best performance on shorted-turn fault detection considering a real dynamic environment subject to voltage unbalance, load variation, and measurement noise.