Abstract: Existing fault diagnosis methods for hydroelectric units either focus on the construction of new fault symptom extraction algorithm based on expert experience, or rely on the machine self-learning algorithm to automatically extract fault symptoms. In order to combine the advantages of the two, this paper proposes a fault diagnosis method for hydropower units based on CNN self-learning and manual experience.The CNN convolutional neural network is used to automatically extract the hidden symptom of hydropower unit. Combined with the artificial symptom in time-domain and frequency-domain commonly used in data analysis of hydropower units. A high-dimensional fusion feature vector is constructed to characterize the vibration characteristics of hydropower units, and the redundant information between features is eliminated by feature screening and dimensionality reduction. Finally, based on XGBoost classification algorithm, a power plant runner room rubbing fault measured data is used to verify that the method can improve the hydroelectric unit fault diagnosis accuracy.