Effectiveness of feature extraction in neural network architectures for novelty detection

This paper examines the performance of seven neural network architectures in classifying and detecting novel events contained within data collected from turbine sensors. Several different multilayer perceptrons were built and trained using backpropagation, conjugate gradient and quasi-Newton training algorithms. In addition, linear networks, radial basis function networks, probabilistic networks and Kohonen self organising feature maps were also built and trained, with the objective of discovering the most appropriate architecture. Because of the large input set involved in practice, feature extraction is examined to reduce the input features, the techniques considered being stepwise linear regression and a genetic algorithm. The results of these experiments have demonstrated an improvement in classification performance for multilayer perceptrons, Kohonen and probabilistic networks, using both genetic algorithms and stepwise linear regression over other architectures considered in this work. In addition, linear regression also performed better than a genetic algorithm for feature extraction. For classification problems involving a clear two class structure we consider a synthesis of stepwise linear regression with any of the architectures listed above to offer demonstrable improvements in performance for important real world tasks.