A Method Of Detecting Gravitational Wave Based On Time-frequency Analysis And Convolutional Neural Networks

This work investigated the detection of gravitational wave (GW) from simulated damped sinusoid signals contaminated with Gaussian noise. We proposed to treat it as a classification problem with one class bearing our special attentions. Two successive steps of the proposed scheme are as following: first, decompose the data using a wavelet packet and represent the GW signal and noise using the derived decomposition coefficients; Second, detect the existence of GW using a convolutional neural network (CNN). To reflect our special attention on searching GW signals, the performance is evaluated using not only the traditional classification accuracy (correct ratio), but also receiver operating characteristic (ROC) curve, and experiments show excelllent performances on both evaluation measures. The generalization of a proposed searching scheme on GW model parameter and possible extensions to other data analysis tasks are crucial for a machine learning based approach. On this aspect, experiments shows that there is no significant difference between GW model parameters on identification performances by our proposed scheme. Therefore, the proposed scheme has excellent generalization and could be used to search for non-trained and un-known GW signals or glitches in the future GW astronomy era.