Abstract: To achieve accurate identification of faults in hydroelectric units, this study first combines the improved complete ensemble empirical mode decomposition with adaptive noise（ICEEMDAN） and correlation analysis methods to decompose and reconstruct the vibration signals. Then, considering the diversity of inputs for convolutional neural networks（CNN）, three different ways of constructing input signals are compared and analyzed for the accuracy of fault diagnosis and recognition by the CNN model. Finally, a square matrix representation of one-dimensional reconstructed signals is selected as the input format for CNN. The effectiveness of this method in improving the accuracy of fault diagnosis in hydroelectric units is validated through real measured data of turbine rubbing faults in a certain power plant.