Gravitational Wave Signal Denoising and Merger Time Prediction By Deep Neural Network
The architecture of our model is delineated as follows. Our model comprises a denoising model (left) and a merger time prediction model (right). In the denoising model, the data first passes through a downsampling network to obtain lowdimensional feature vectors. A separator then analyzes the deep relational patterns within these vectors to extract GW features embedded therein. An upsampling network is responsible for utilizing features from each layer of the dimensionality reduction network, through trimming and concatenation, to ultimately reconstruct the waveform. In the merger time prediction model, we reuse the downsampling structure in the denoising model and finally use two linear layers to obtain the merger time.
