Diagnoses for machine vibrations based on self-organization neural network

The authors present a diagnostic system for the recirculating system of an online nuclear power station, based on the neural network. In the learning phase, signals of sensors which monitor the mechanical operation of the recirculating system are preprocessed mainly by spectrum analysis to produce sets of spatiotemporal patterns. The neural network serves the kernel of the diagnostic system, whose main function is the self-organization of a feature map of these spatiotemporal patterns. Each spatiotemporal pattern is a high-dimensional vector composed of 128 elements, and, after self-organization, a reduced 2D feature map is established as composite diagnostic panel (CDP). In the detecting phase, all monitoring signals are fed into the CDP. The basic concept is to make the sensor signals visible. Computer simulations showed that the idea is a suitable approach to the requirement for ensuring visibility of these signals