深度学习语义分割方法用于双致密天体并合引力波数据处理研究OA北大核心CSTPCD

Recent studies have shown that deep learning(DL)based gravitational wave(GW)search holds the promise of addressing the inefficiencies of matched filtering method.Compared to matched filtering for GW search and Bayesian posterior-based or deep learning-based methods for GW source parameter estimation,the DL based end-to-end GW search lacks the time-frequency information of the source,such as the arrival time and the duration of the signals within the sensitivity range of the detectors.This study represents the first implementation of fine-grained classification of time-frequency points in strain signals and explores a gravitational wave search approach based on image semantic segmentation.The data is synthesized by combining real noise detected by the Hanford interferometer and simulated physical signals.It is then transformed into time-frequency images using the Q-transform,and a pixel-level annotated dataset is constructed.The time-frequency image semantic segmentation model is built and trained,and the results demonstrate the feasibility of image semantic segmentation methods in the analysis of gravitational wave signal data.The image semantic segmentation methods have the potential to become a component of the gravitational wave signal search pipeline for further extraction of time-frequency information of signals in the future.