Unsupervised Learning Approach to Anomaly Detection in Gravitational Wave Data

TL;DR

This work tackles unsupervised anomaly detection in gravitational-wave time-series by training a variational autoencoder (VAE) on noise-only data to learn normal detector noise patterns. Anomalies, such as gravitational-wave signals, are identified via spikes in the reconstruction loss $||x-\hat{x}||^2$, with a temporal LSTM-based VAE architecture designed to capture noise dynamics. On LIGO H1/L1 data, the method achieves a ROC-AUC of 0.89 and a F1 score of 0.857 when distinguishing GW events from noise, outperforming a vanilla autoencoder. The approach offers a scalable, template-free framework for detecting both known and potentially new phenomena in gravitational-wave data and can be extended to other time-series domains in physics.

Abstract

Gravitational waves (GW), predicted by Einstein's General Theory of Relativity, provide a powerful probe of astrophysical phenomena and fundamental physics. In this work, we propose an unsupervised anomaly detection method using variational autoencoders (VAEs) to analyze GW time-series data. By training on noise-only data, the VAE accurately reconstructs noise inputs while failing to reconstruct anomalies, such as GW signals, which results in measurable spikes in the reconstruction error. The method was applied to data from the LIGO H1 and L1 detectors. Evaluation on testing datasets containing both noise and GW events demonstrated reliable detection, achieving an area under the ROC curve (AUC) of 0.89. This study introduces VAEs as a robust, unsupervised approach for identifying anomalies in GW data, which offers a scalable framework for detecting known and potentially new phenomena in physics.

Figures (6)

• Figure 1: Simplified Diagram of the LIGO detector abbott2016observation. The detection mechanism is explained in \ref{['ligo-sec']}. Subfigure (a) shows the locations of the L1, and H1 detectors used for our analysis.
• Figure 2: Strain data of event GW150914. The GW signal occurs at time = $16.5$ s. The data is whitened and band-passed in the range $[20,400]$.
• Figure 3: Toy example on autoencoders: After training, the autoencoder reconstructs inputs effectively when the test data shares the same distribution as the training data. However, in the presence of anomalies (red), the autoencoder fails to reconstruct them accurately, resulting in a spike in the reconstruction error at the anomaly's occurrence.
• Figure 4: Illustration of our VAE model with LSTM units (not to scale). The encoder is the lower left block and the decoder is the upper right block. The objective \ref{['objective']} is maximized by back-propagating it through the network to find the optimal parameters $\theta,\phi$.
• Figure 5: Performance on data from Event GW150914: quadratic loss on the stream of data. Notice how it peaks in the presence of the gravitational wave.