Core-collapse supernovae are potential sources of gravitational waves that could be detected by current and future detectors, and their detection and analysis are of great importance for understanding the explosion mechanism. Since matched filtering cannot be used for these signals due to the stochastic nature of the waveforms, detection methods based on time-frequency representation have been developed. Recently, deep learning has been applied to the analysis of gravitational wave data and has the potential to greatly improve our ability to detect and analyze these signals.
In this study, we apply a convolutional neural network to detect and classify gravitational waves from core-collapse supernovae. The model is trained on waveforms obtained from 3D numerical simulations, injected in real noise of O3 observing run. We also apply class activation mapping technique to visualize from which part of the input the model predicted the result. The results show that our model is able to classify 9 different waveforms and noise with 96.9% accuracy at 1 kpc. The maps visualized by class activation mapping technique show that the model's predictions are based on g-mode shapes of input spectrograms.