In this article, we are interested in improving the prediction of the remaining useful operating time of a complex system whose state is represented by multivariate time series. We present and evaluate two approaches for measuring the improvement of the Remaining Useful Life (RUL) prediction using four different state-of-the-art machine learning approaches based on deep learning. The first method that we propose is based on re-sampling the training data set in order to reduce the errors locally. The second proposed method relies on automatically detecting and using break-points in the signals to improve the training step. We show that break-point detection techniques allow a significant improvement of the RUL prediction performance with gains of more than 27% on the mean absolute error (MAE) regardless of the neural architecture used, which demonstrates the genericity of our approach.