Research On Gravitational Wave Data Analysis Based On Convolutional Neural Network

With the continuous improvement of LIGO / Virgo sensitivity,more and more gravitational wave events have been detected.Matched filtering method is an important data processing method for gravitational wave detection,but it requires a large number of matching templates and low computational efficiency,and cannot realize real-time detection.Therefore,it is urgent to find a method that can analyze gravitational wave signals quickly and in real time.With the application of deep learning technology in image processing,medical diagnosis,unmanned driving and other fields.Experts in the field of astronomy also try to apply deep learning technology to gravitational wave signal processing.Compared with the matched filtering method,the computational efficiency of deep learning is greatly improved,and it is expected to be applied to the real-time detection of gravitational wave signals.I have studied the application of deep learning in gravitational wave detection.The main work contents are as follows:(1)The structural complexity of convolutional neural network(CNN)and its performance in gravitational wave detection are studied.The deep learning gravitational wave detection models with different convolution kernel size,the number of convolution kernels(the width of the model)and the number of convolution layers(the depth of the model)are discussed.In addition,we studied the detection effect of the model with batch normalization(BN)before the full connection layer,and found that the detection accuracy of the model with single convolution layer increased from about 50% to more than 90%.The results of this work provide a potential new method for the compression of the number of matched filter templates.After matched filtering,the number of matched templates may be greatly reduced through BN layer and full connection layer.(2)The response of convolutional neural network(CNN)gravitational wave detection model to glitch signal is studied.LIGO strain signal includes not only stationary Gaussian noise,but also nonstationary and non-Gaussian glitch components.Whether the glitch signal can be misjudged as gravitational wave signal by the gravitational wave detection model trained with stationary Gaussian noise as background noise is studied.The responses of the CNNs to the sine-Gaussian,Gaussian,ring-down,scattered light-like,and whistle-like glitches are studied.We find that the CNNs have strong generalization ability for sineGaussian,Gaussian,and whistle-like glitches,whose false alarm probabilities(FAPs)are less than 2%,in contrast to the case of the ring-down and the scattered light-like glitches,in which the FAPs are far larger than those of the stationary Gaussian noises.The FAP of ringdown glitches is about 4%–8%,and of scattered light-like glitches is about 22%–34%.The responses of the sine-Gaussian and the ring-down glitches with different central frequencies in the CNNs are also studied.We find that when the center frequency of glitches falls within the main frequency range of the trained GW signals,the FAPs of the glitches will be much larger than those of the stationary Gaussian noises.The probability of the glitches being misjudged as GW signals may even exceed 30%.