New methods to assess and improve LIGO detector duty cycle
A network of three or more gravitational wave detectors simultaneously taking data is required to generate a well-localized sky map for gravitational wave sources, such as GW170817. Local seismic disturbances often cause the LIGO and Virgo detectors to lose light resonance in one or more of their component optic cavities, and the affected detector is unable to take data until resonance is recovered. In this paper, we use machine learning techniques to gain insight into the predictive behavior of the LIGO detector optic cavities during the second LIGO-Virgo observing run. We identify a minimal set of optic cavity control signals and data features which capture interferometer behavior leading to a loss of light resonance, or lockloss. We use these channels to accurately distinguish between lockloss events and quiet interferometer operating times via both supervised and unsupervised machine learning methods. This analysis yields new insights into how components of the LIGO detectors contribute to lockloss events, which could inform detector commissioning efforts to mitigate the associated loss of uptime. Particularly, we find that the state of the component optical cavities is a better predictor of loss of lock than ground motion trends. We report prediction accuracies of 98% for times just prior to lock loss, and 90% for times up to 30 seconds prior to lockloss, which shows promise for this method to be applied in near-real time to trigger preventative detector state changes. This method can be extended to target other auxiliary subsystems or times of interest, such as transient noise or loss in detector sensitivity. Application of these techniques during the third LIGO-Virgo observing run and beyond would maximize the potential of the global detector network for multi-messenger astronomy with gravitational waves.