Artificicial (i.e. Artificial) neural networks and their applications in diagnostics of incipient faults in rotating machinery

Abstract

In an effort to curtail rising operating costs, machinery condition monitoring and diagnostics are being increasingly used as part of predictive maintenance programs. Vibration analysis is currently among the most effective tools in machinery condition monitoring and diagnostics but has proven difficult to automat fully. Artificial neural networks, patterned after neurological systems, provide a heuristic, data based approach to problems and have demonstrated robust behavior when faced with unique and noisy data. Thus neural networks may provide an alternative or complement to conventional rule based expert systems in machinery diagnostics applications. Research is presented wherein a series of neural networks utilizing the highly successful backpropagation paradigm are configured to provide machinery diagnostics for comparatively uncomplicated mechanical systems. Through observation of their responses to minor architectural changes and performance upon presentation of genuine and artificially generated vibration data, an effort is made to ascertain their utility in more complicated systems.