Abstract: To accurately predict the specific failure modes that may occur in the future of equipment, a device failure trend prediction method based on double deep learning model is proposed. Wavelet packet transform (WPT) was used to extract the time-frequency domain features of the vibration sensor signal. The designed GAM-BiLSTM-RF model was used to predict the future operation trend of the equipment, and obtain the time series data corresponding to the future operation trend. The designed GAM-BiLSTM-ET model was used to extract the deep features of the future operation trend data of equipment, and judge the type and severity of equipment faults according to the extracted features. The experimental results on the open equipment fault datasets (IMS and XJTU-SY) show that the proposed two-step prediction method can accurately predict the future operation trend of equipment, and effectively improve the accuracy of equipment fault prediction. In addition, the performance comparison with some baseline models further verifies that the proposed method is effective and stable.