DETECTING LONG-TERM TRENDS IN TURBO-GENERATOR STATOR END-WINDING VIBRATIONS THROUGH NEURAL NETWORK MODELLING

The accurate assessment of remaining useful life based on condition monitoring variables is not a trivial task, since long-term trends are often obscured by short-term fluctuations. Short-term variations in such variables also tend to overshadow the long-term drift in magnitude. Stator end-winding vibrations are one of the key indicators of the remaining useful life of turbo-driven generators. In this paper, a technique is developed to separate long-term drifts in stator end-winding vibrations from short-term fluctuations. The technique rests on the fact that short-term variations in winding vibrations are largely affected by operational variables measured on a turbo generator, including load and temperature. These dependencies can be captured in a model reflecting the short-term behaviour of the vibration amplitudes. The long-term trend in vibration amplitude is, however, not governed by the same relationships. It is hence possible to extract the long-term trend from the overall behaviour by subtracting the short-term effects of operational variables from the overall behaviour. In this way, a reliable long-term trend is obtained, from which remaining life assessments could be made.