Gravitational wave search by time-scale-recursive denoising and matched filtering

In our previous work [Physical Review D, 2024, 109(4): 043009], we introduced MSNRnet, a framework integrating deep learning and matched filtering methods for gravitational wave (GW) detection. Compared with end-to-end classification methods, MSNRnet is physically interpretable. Multiple denoising models and astrophysical discrimination models corresponding to different parameter space were operated independently for the template prediction and selection. But the MSNRnet has a lot of computational redundancy. In this study, we propose a new framework for template prediction, which significantly improves our previous method. The new framework consists of the recursive application of denoising models and waveform classification models, which solve the problem of computational redundancy. The waveform classification network categorizes the denoised output based on the signal’s time scale. To enhance the denoising performance for long-time-scale data, we upgrade the denoising model by incorporating Transformer and ResNet modules. Furthermore, we introduce a novel training approach that allows for the simultaneous training of the denoising network and waveform classification network, eliminating the need for manual annotation of the waveform dataset required in our previous method. Real-data analysis results demonstrate that our new method decreases the false alarm rate by approximately 25%, boosts the detection rate by roughly 5%, and slashes the computational cost by around 90%. The new method holds potential for future application in online GW data processing.