Omniformer: A real time transient noise source localizer
ORAL
Abstract
Transient noise is characterized by short-lived bursts of interference from many sources, such as seismic, acoustic, and electronic noise. It poses a significant challenge in gravitational wave detection. This noise often renders interferometer data unusable, necessitating its exclusion and reducing sensitivity. To address this, auxiliary channels are used. These are independent time-series data streams from sensors monitoring environmental, electromagnetic, seismic and many other disturbances within an interferometer. They help flag noise-contaminated time segments.
By cross-referencing these channels with the primary strain data, segments are categorized based on their signal-to-noise ratio (Signal here referring to the changes in strain channel due to astrophysics events such as CBCs, BH-BH mergers etc. Noise, is the portion of the time series where the origin of noise is non-astrophysical and is validated by coincident bursts in both strain and auxillary channels indicating external interference)
We propose Omniformer, a real-time noise localization model leveraging a pretrained transformer architecture. The model first ingests the full observation dataset, which includes main strain and auxiliary channel data streams. In its initial step, Omniformer processes these inputs and monitors for temporal segments with correlated noise patterns between the main strain and auxiliary channels. When correlated noise is detected, the model assigns a probability score to each flagged segment, indicating the likelihood of transient noise presence. Using these scores, Omniformer then localizes the origin of the transient noise within the interferometer. The model is trained on previous observation runs, enabling it to identify temporal patterns and flag relevant events in real time.
By cross-referencing these channels with the primary strain data, segments are categorized based on their signal-to-noise ratio (Signal here referring to the changes in strain channel due to astrophysics events such as CBCs, BH-BH mergers etc. Noise, is the portion of the time series where the origin of noise is non-astrophysical and is validated by coincident bursts in both strain and auxillary channels indicating external interference)
We propose Omniformer, a real-time noise localization model leveraging a pretrained transformer architecture. The model first ingests the full observation dataset, which includes main strain and auxiliary channel data streams. In its initial step, Omniformer processes these inputs and monitors for temporal segments with correlated noise patterns between the main strain and auxiliary channels. When correlated noise is detected, the model assigns a probability score to each flagged segment, indicating the likelihood of transient noise presence. Using these scores, Omniformer then localizes the origin of the transient noise within the interferometer. The model is trained on previous observation runs, enabling it to identify temporal patterns and flag relevant events in real time.
–
Presenters
-
Shantanusinh Parmar
- Marwadi University