Abstract: When a hydroelectric unit operates under non-steady working and abnormal conditions, it will produce violent vibration and harsh noises.In order to ensure the safe and reliable operation of a unit, a bulb tubular hydropower unit is taken as the research object, the vibration and noise of each unit part were monitored in real time by arranging high-precision acceleration and audio sensors, and multi-source fusion signals of vibration and audio were collected.The Kernel Principal Component Analysis(KPCA) and the improved K-Means clustering algorithm are used to extract the root mean square parameter of the frequency and amplitude of the multi-source fusion signal, and the energy distribution, eigenvalues of faults such as the turbine blade collision, body knock and generator partial discharge are obtained.Based on the energy distribution and eigenvalues, a six-dimensional eigenvector that can reflect the state of a unit is constructed.Combined with the on-site fault simulation test, the corresponding fault can be accurately identified by the extraction method.The research results can provide strong support for the maintenance of the units.