• convolutional autoencoder
• glitches in gravitational wave detectors
• gravitational wave detector characterization
• transient noise in gravitational wave detectors
• unsupervised clustering

Gravitational waves interferometers are complex and sensitive detectors, whose data are non stationary and non Gaussian. Short duration disturbances, called ’glitches’, are caused by the instrument itself or by its interactions with the environment. These transient noises are particularly concerning as they can mimic gravitational wave signals, and have both high rate and high signal-to-noise ratio. Great effort has been made in the latest years to understand their causes and mitigate them: in particular as glitches differ significantly in terms of duration and frequency, their classification is crucial to trace back their origin. Due to glitches complexity, huge number and time evolving nature, machine learning techniques find great application in this field. This thesis proposes an unsupervised clustering algorithm able to group transient noise into different classes according to their morphology in a time-frequency map, using a neural network called ’autoencoder’ and a density-based clustering algorithm, without any prior knowledge on the data it is applied to. This method is successfully tested on LIGO Hanford detector glitches, and applied to latest Virgo data, clustering one week, randomly selected, of Virgo transient noise and contributing to a candidate event validation. The information acquired could enhance detector characterization and transient noise mitigation, improving gravitational-wave searches