Abstract: The vibration signals of the hydropower unit contain a wealth of unit status information, and if the fault characteristics contained in them can be fully extracted and effectively utilized, it will bring great convenience to identify the unit status and diagnose the unit fault. In order to fully extract the fault features contained in the vibration signals, the Deep Residual Shrinkage Network（DRSN） is combined with Empirical Mode Decomposition（EMD）, the former mining the hidden information of the data, and the latter extracting the time-frequency features, and then forming the fusion features. Subsequently, in order to make effective use of these fault characteristics, the Improved Light Spectrum Optimizer（ILSO） was used to optimize the kernel function parameter G and penalty coefficient C of the Support Vector Machine（SVM） to improve the classification accuracy of SVM. The analysis shows that the method can deepen the excavation of the fault characteristics of hydropower units to a certain extent, and improve the efficiency and accuracy of fault diagnosis.