Maintenance activities can be better planned by employing machine learning technologies to monitor an asset’s health conditions. However, the variety of observable measures (e.g. temperature, vibration) and behaviours characterizing the health degradation process results in time-consuming manual feature extraction to ensure accurate degradation stage recognitions. Indeed, approaches able to provide automatic feature extraction from multiple and heterogeneous sources are more and more required in the field of predictive maintenance. This issue can be addressed in a data-driven fashion by using feature learning technology, enabling the transformation of minimally processed time series into informative features. Given its capability of discovering meaningful patterns in data while enabling data fusion, many feature learning approaches are based on deep learning technology (e.g. autoencoders). In this work, an architecture based on autoencoders is used to automatically extract degradation-representative features from minimally preprocessed time series of vibration and temperature data. Different autoencoder architectures are implemented to compare different data fusion strategies. The proposed approach is tested considering both the recognition performances and the quality of the learned features with a publicly available real-world dataset about bearings’ progressive degradation. The proposed approach is also compared against manual feature extraction and the state-of-the-art technology in feature learning.