A methodology for the analysis and monitoring of vibration data using artificial neural systems

This dissertation deals with the use of artificial neural networks for the monitoring and diagnosis of components in nuclear power plants. The technology of neural networks provides an attractive complement to traditional vibration analysis because of its potential to operate in real-time and to handle data which may be distorted or noisy. The technique enhances traditional vibration analysis and provides a means of automating the monitoring and diagnosis of vibrating equipment.; The study is conducted in two phases. First, the mechanical behavior of rotating machinery is studied, and a neural network system is developed for the detection of faults in rolling element bearings (REB). The system learns association between features in the spectrum and operating states of the bearings, the system is tested using data from an aging simulation bench test.; A technique is presented for modelling the relationship among sensors in a machine that is shown to be very effective for identifying changes in operating states. REB's are especially interesting components because they are responsible for a large fraction of the malfunctions in manufacturing equipment. In an average nuclear power plant about one third of all measuring points is allocated to REB's. The second phase of the study is related to the problem of the vibration of the internals of pressurized water reactors (PWR). The vibratory behavior of the internals in a PWR can be identified and monitored using ex-core neutron data from ionization chambers located outside the vessel. The data collected from these detectors provide information regarding the behavior of mechanical components, the presence of contacts between the core barrel and the pressure vessel, and more importantly, a means of verifying the integrity of components in the system.; A neural network-based methodology is described for identifying the vibration mode of the core barrel, and for detecting a particular family of mechanical failures. Features are extracted from the neutron noise spectra, and used for training neural network models to identify the different states of vibration typically exhibited by PWR's. The technique was tested on data collected over 93 fuel cycles from twenty-eight 900 MWe pressurized water reactors in France.