Modelling complex processes from raw time series increases the necessity to build Deep Learning (DL) architectures that can manage this type of data structure. However, as DL models become deeper, larger and more diverse datasets are necessary and knowledge extraction will become more difficult. In an attempt to sidestep these issues, in this paper a methodology based on two main steps is presented, the first being to increase size and diversity of time-series datasets for training, and the second to retrieve knowledge from the obtained model. This methodology is compared with other approaches reported in the literature and is tested under two configuration setups of Condition-Based Maintenance problems: fault diagnosis of bearing, and fault severity assessment of a helical gearbox, obtaining not only a performance improvement in comparison, but also in retrieving knowledge about how the signals are being classified.