Deep learning algorithms have attracted attention due to their powerful time series forecasting capabilities and the advantages of being able to process massive samples of massive data in real time. Aiming at the problems of low accuracy， missing diagnosis， and difficult prediction in the vibration fault diagnosis of hydraulic turbine systems， a hydraulic turbine system fault prediction method based on deep learning long short time memory （LSTM） networks combined with deep belief networks （DBN） is proposed. This method combines wavelet packet energy bands with time-frequency domain index information to extract high-dimensional fault statistical features， and uses the adaptive feature extraction capabilities of the DBN deep network to perform high-dimensional feature representations on the original fault data， to more accurately determine the types of faults， and to predict the possible vibration faults of the system in the future with the powerful predictive ability of LSTM on time series signals. The effectiveness of the algorithm is verified by engineering experiments.