Hydroelectric-generator unit fault early warning method based on distribution estimation

In order to realize hydroelectric-generator's auto fault early warming, a new method based on distribution estimation is presented. Contrasting to other classical methods, the training data this method used are not fault data, but history data under normal operation. Generator's vibration is looked as independently identical distribution observation samples which drawn from an underlying probability distribution and with this assumption generator's vibration pattern can be carried out by using SchÖlkopf's one class support vector machine method. Then with this pattern fault early warming can be carried out when new observation comes. This method has the property of dealing observation directly without complicated pretreatment, rapidity, and simplicity, it also has adaptive capacity to the data missing and change of operation parameters. Simulation on real observation shows its ability of learning generator's normal vibration pattern and early warning fault pattern. Distribution estimation provides a new way for fault early warming.