Gravitational wave signal extraction against non-stationary instrumental noises with deep neural network
The architecture of our denoising autoencoder is delineated as follows. The input data is first subjected to normalization and segmentation (top left), resulting in the formation of overlapping subsequences. Each subsequence is then processed through the encoder to extract its characteristic feature vector. This is followed by passage through three bidirectional LSTM layers to yield predictive values. The final reconstructed waveform is then attained in the Dense layer (top right).
