Gravitational_wave_detection.pptx
- 1. Gravitational wave detection and denoising algorithm Hook There are very extremely small waves in the universe. The waves are full of noise. how do i find it？ Key details Gravitational wave noise is smaller than the noise of the equipment that detects it, so it is necessary to classify the noise. Noise can be classified into about 22 types using software. Visualize the waveform data and apply a layer to each noise data. Let the machine learning supervised and unsupervised learning do the work in general against the example of image averaging. Supervised learning can efficiently perform image processing, which is a heavy workload. Unsupervised learning allows the computer to do the method itself, which is slower, but ensures new possibilities, such as finding new classifications. Merge it so that the 22 images are as similar as possible. More details The proposed unsupervised learning method consists of two architectures: a variational autoencoder (VAE) and invariant information clustering (IIC). The VAE is used to learn the features from the time–frequency spectrogram (2D images) of transient noise, and the IIC classifies the transient noise from the features that are learned by the encoder of the VAE. Before we present the details of the method, we explain the target dataset. Target dataset The Gravity Spy dataset16, which is the input dataset, is an image set of transient noise obtained from the LIGO O14. Omicron software19 searches for transient noise in time-series data, and Omega Scan20 software generates an image of the time–frequency spectrogram of Each transient noise using Q-transformation20,38. Q-transformation is a method that estimates the frequency component of the time-series data by setting a window function on each time–frequency component, generating a 2D image of the time–frequency spectrogram. The spectrogram image of each transient noise in the Gravity Spy dataset has four time durations (0.5, 1.0, 2.0, and 4.0 s) at the center,. are related to cause. Pre-processing The pre-processing applied to the Gravity Spy dataset for the training of our proposed architecture. Considering the characteristics of the time–frequency spectrogram, a small displacement in the time direction does not change its physical characteristics because this operation can be interpreted as a change in the event time. of transient noise, and it makes the architecture realize the classification of transient noise that does not depend on small displacements in the time direction. Conversely, a possible small displacement of the spectrogram in the frequency direction changes its physical characteristics. shifted images fall into different classes to that of the original image in the classification.