Using normal to find abnormal: AI-based anomaly detection in gravitational wave data

Abstract

The detection and classification of anomalies in gravitational wave data plays a critical role in improving the sensitivity of searches for signals of astrophysical origins. We present ABNORMAL (AI Based Nonstationarity Observer for Resectioning and Marking AnomaLies), a deep neural network (DNN) model for anomaly detection that is trained exclusively on simulated Gaussian noise. By removing dependence on real data for training, the method resolves a circular paradox in anomaly detection: training on real data implicitly involves prior segregation of stationary from non-stationary data but this is not possible unless all anomalies are detected first. ABNORMAL is an autoencoder-based DNN, commonly used in anomaly detection, with the key innovation that it is trained to predict statistical features of noise rather than reconstructing the noise time series themselves. The statistical features are obtained by applying Gabor and Wavelet filter banks, which implement time-frequency analysis, and are subsequently combined through multi-view fusion using a dual-path architecture. We quantify the performance of our method on simulated and real LIGO data. Application to data from the O1 to O3b observational runs uncovers a rich landscape of anomalies over different timescales, including many that do not fit within known classes.