Detecting Gravitational Waves as Anomalies with Convolutional Autoencoders

In this chapter, we summarize a generic approach to the analysis of gravitational wave (GW) data: we consider the GW to be the anomaly in the signal received by the detector and focus on detecting such anomalies. This approach works because detectable gravitational waves are rare (at the moment), transient signals, distinct from the slow-varying, non-stationary noise typically present in the detectors. The anomalies under investigation arise mainly from gravitational waves produced by the merger of binary black hole systems. However, our analysis extends beyond GW signals to encompass glitches identified within the real LIGO/Virgo dataset, accessible through the Gravitational Waves Open Science Center. Our anomaly detection process is based on deep learning techniques, specifically convolutional autoencoders trained on both simulated and real detector data. We demonstrate the efficacy of our method by reconstructing injected GW signals and explore how the detection of anomalies is influenced by the strength of the gravitational wave, quantified via the matched filter Signal-to-Noise Ratio (SNR). Furthermore, we apply our methodology to localize anomalies within the temporal domain of the time-series data that models the gravitational wave. The validity of our approach is confirmed by applying it to real-world data containing verified gravitational wave detections: our method is able to generalize and identify GW events not included in the training dataset.