Abstract Gravitational wave (GW) analysis is attracting widespread attention as an emerging research field. As the presence of substantial noise in GW signals, and the characteristics of inspiral and merger stage are different, coupled with the sidelobe effect caused by window length, traditional time-frequency analysis methods face significant challenges in accurately analyzing the frequency variations of GW signals. This poses a major limitation in the precise analysis stage following GW detection. Therefore, we proposed a novel method of Seasonal-Trend decomposition using Loess with Multilayer Perceptron (STLMLP), for predicting and validating the accuracy and effectiveness of GW frequency variations. Experiment results on three noiseless GW templates demonstrate that STLMLP exhibits the adaptability and highest prediction accuracy for the dynamic frequency variations of GW signals compared to five state-of-the-art machine learning and deep learning methods. Furthermore, experiments conducted on three noisy actual GW data compared with the state-of-the-art method of Fourier-based synchrosqueezing transform (FSST) in the signal processing domain confirm that STLMLP maintains lower error in predicting frequency change over the whole duration of the actual noisy GW signals.
Read full abstract