Unsupervised Classification of Transient Noise in Gravitational Wave Observation

Abstract

Gravitational wave (GW) predicted by General Relativity has recently become observable. In gravitational wave observation, non-stationary and non-Gaussian noise, called ``transient noise'', frequently appear. It is known that the transient noise might cause the instabilities in the detector. Transient noise might also mimic and obscure the GW signal. Identifying and classifying transient noise have a possibility of improving the detector performance. This study employed the deep learning (variational auto-encoder) to extract latent variables of transient noise. A tool that visualizes the latent variables embedded in a 3D space by using UMAP was developed. Because the tool can also display the corresponding input images, the visualization tool enables to the analysis of noise distribution in the latent space alongside the input images. Unsupervised classification was also performed on the transient noise in the latent space. By evaluating clustering instability and misclassification rate, the suitable class numbers were estimated, and the classification results were discussed.