Abstract: Accurate prediction from operation and maintenance personnel of the upcoming disk failure is the key to ensure data security. However, unbalanced data and inaccurate disk characteristic marking affect the accuracy of prediction. This paper proposes a disk failure prediction method based on pre_Failure Reseting Window (pre_FRW) data processing and combining convolutional neural network (CNN) and long short-term memory network (LSTM), namely pre_FRW-CNN-LSTM. The pre_FRW data processing could not only solve the sample imbalance, but also reduce the potential fuzzy samples. The CNN-LSTM model structure could extract the spatial characteristics of the data, and it could also effectively capture the dependencies between time series. Experiments on real monitoring data sets show that the disk failure prediction method of pre_FRW-CNN-LSTM improves the failure prediction rate by 2%-10% compared with other methods in the industry and maintains a low false alarm rate.