Hydroelectric unit voice signals contain a significant amount of valuable information reflecting their internal mechanical state. To accurately extract the voiceprint features of rubbing faults in hydroelectric units

this paper presents a hydroelectric unit rubbing fault voiceprint recognition model based on the fusion of Ensemble Empirical Mode Decomposition (EEMD) and Convolutional Neural Network (CNN). First

we use EEMD to decompose a noise signal from a hydroelectric unit into several Intrinsic Mode Functions (IMFs) and a residue component (Res); we use these IMFs and Res

along with the original signal

to construct a fusion feature vector. Then

the vector is used as an input to train a CNN deep learning neural network

with the normal and rubbing fault test data as samples

so as to obtain a rubbing fault recognizer for hydroelectric units. This new method is validated against the rubbing test data from both the hydro-mechanical coupling test stand and the in-situ experiment

with an average accuracy of 99.8%

demonstrating its performance superior to other recognition models for the rubbing faults of hydroelectric units.