Neural networks in vibration analysis of gas turbines

Artificial neural network technology was applied to the problem of vibration analysis of a large gas turbine engine. The subject domain of gas turbine vibration was researched and neural network input features describing the turbine operation were developed. Different neural network architectures were researched and applied to the vibration analysis. Both supervised and unsupervised neural networks were created and tested, with an emphasis on the unsupervised Fuzzy ART network. A Fuzzy ART network capable ot analyzing a 32768-point vibration spectrum and detecting 96% of changes introduced throughout the spectrum was developed. The theory of resonance fields and the technique of feature separation using a priori information were developed and utilized in the research. Individual case-vigilance techniques were formalized and used to reduce noise clutter in the processing of the spectrum for neural net input. Through training, the neural network automatically created vibration amplitude acceptance envelopes, or classification regions, through neural compression of the spectrum, to perform continuous monitoring of changes in individual turbine-component vibration. This neural network has application to detection of narrowband vibration trends, important for condition-based maintenance of turbomachinery. In the course of this research, turbine data sets were recorded and digitized from LM2500 and LM6000 gas turbine engines. Descriptions of the multi-channel data acquisition and hybrid analog/digital anti-aliasing filter system are provided. A method to incorporate a priori information into the network training set was developed. With knowledge of the network training operation, different sections of the input space could be separated to control the neural classification of those sections.