Remaining useful life (RUL) prediction is a key technology to realize intelligent health management and maintenance of new generation aeronautic equipment. However, traditional data-driven methods are difficult to fully extract the performance degradation characteristics of complex systems, resulting in large lags in RUL prediction results, which is harmful to engineering applications. Aiming at the above problems, an end-to-end improved LSTM autoencoder model (LSTM-DAE) is proposed, which improves the model’s ability to extract performance degradation features of complex systems by introducing multiple fully connected layers. Without any prior degenerate knowledge, the original high-dimensional observation sequence can be directly mapped to a one-dimensional performance degradation curve based on LSTM-DAE. Then the curve can be matched by similarity based on Euclidean distance to predict RUL. The effectiveness of the model is verified on the aero-engine public dataset, and the results show that the distribution of RUL estimated values is more concentrated. Compared with existing method, the Score metric of the LSTM-DAE is reduced by 9%. The probability of lagged prediction is significantly reduced. It means that the reliability of life prediction is further improved.