The application of machine learning in geophysics has shown to be effective in automating various supervised tasks, such as earthquake detection, localisation, classification and denoising. A large majority of these approaches are based on supervised learning, where the target task can be quantified using a well-defined loss function. In this study, we present a framework that exploits unsupervised techniques to acquire knowledge from continuous geophysical time series, based on a limited priori on the data. Using clustering and dimensionality reduction, we summarise large amounts of information into a low-dimensional representation to facilitate pattern recognition. We apply this framework to various seismic datasets, demonstrating its ability to describe data content at different time and frequency scales. In addition, we investigate the ability of our approach to identify mutual information between seismic data and external data such as GNSS and meteorological data to better constrain the understanding of the physical origin of different signatures, with applications to the Mexican subduction zone over a decade of continuous data including slow slip events.