Gas turbine vibration analysis with fuzzy ART neural network

Abstract

High-resolution turbine spectral data was analyzed using a fuzzy ART neural network. The network was configured as a novelty detector to automatically detect changes in the turbine operating characteristics as evidenced in the vibration spectrum. To accomplish reliable novelty detection of high-resolution spectral data, the characteristics of fuzzy ART with regards to prototype hyper-dimensions and the hyper-dimensional areas around these prototypes that could cause adaptation of the prototype were analyzed and a theory of resonance fields constructed. Using resonance field theory, a fractional dimensional offset was created to allow definite separation of closely spaced features in the input spectrum. A helpful side effect of the introduction of fractional-dimension offsets was a significant increase in speed of learning of the high-resolution data. The use of individual-case based vigilance allowed the 32768-point, 48 dB range spectrum to be learned in 228 neurons to a controlled 5% of full-scale accuracy, at real-time speeds for this application.