Dictionary learning with application to vibration analysis

This thesis aims to apply dictionary learning techniques to a real-world problem: audio signals recorded from a machine tool for high-precision engravings. We want to extract information about the working stages and possible failures using only microphone data as input and dictionary learning, without the need to build complex mathematical-physics models. The use of DL methods allows for building a flexible model tailored to the data (similarly to other machine learning techniques) while retaining human interpretability (the dictionary contains the building "atoms" whose sparse linear combinations can supposedly represent the input). This can be useful for sparse recovery, compressing, denoising, etc., or (as is our case) classification. The Short Time Fourier Transform (STFT) of the signal already provides useful insights into the process; is DL able to recognize the most relevant features?