Deep Learning For High Frequency Gravitational Wave Simulation Detection Data Processing

Gravitational waves(GWs)in low-and intermediate frequency bands were first detected in 2016,which opens a new era of astronomical observations.Therefore,high frequency gravitational waves(HFGW)detection must play an important role in future.Given the detectable signals of HFGWs are significantly weak,the sensitivity of the detector is expected to be very high.This implies that the amount of detected data should significanltly big.Therefore,it is particularly desired to develop the relevant methods to quickly and accurately process the big detected data.Although the matched filtering method has been successfully used to process the experimental detection data by the LIGO-Virgo group for the GWs in lowand intermediate frequency bands,the huge computational cost is required practically.Therefore,it is particualy desired to develop the low-cost methods for processing the big detection data of the HFGWs in future.In recent years,deep learning has played the more and more important roles in the big data processing,even for the detections of the low-and intermediate frequency gravitational waves.It has shown that,compared with traditional matched filtering technology,deep learning can not only accurately estimate the physicaly information of the detected GWS,but also greatly reduce the computing time.Therefore,it is desireable to develop the similar deep learning method to process the detected data for the future HFGW detection.Based on the convolutional neural network(CNN)and template matching,in this thesis we focus on the data process of the simulated detection of the HFGWs.First,basing on the CNN model proposed first in Phys.Rev.D 97,044039(2018),wherein the experimental detection data of the intermediate frequency gravitational waves was accurately classified,we develop a CNN classification model for the spatial-domain simulation detection data of the HFGWs.By adding random dropout layers and L2 regularization and other optimization measures,we show that the potential overfitting phenomenon can be effectively alleviated.By using the optimized CNN model developed above,we have implement the train and test of the simulated time-domain detection dataset of the HFGWs.It is shown that,on the testset the area under the receiver operating characteristic curve(ROC)is 0.986,which indicates that the accuracy of the model is sufficiently high.Furthermore,the frequency and amplitude parameters of the simulatively detected HFGWs were estimated from the timedomain simulated detection data by using the usual template matching.The estimated parameters have a narrow confidence interval under the 95% confidence level,which indicate that the relevant estimateion is very reliable.The robustness of the CNN model and template matching has been verified by using multiple simulated detection datasets.Finally,aiming at the detection of the stochastic GWs,we proposed an effective method by using the cross correlation to identify the time-domain simulated detection data of the stochastic HFGWs.The simulated data are generated by the synchronous and remote simulated detections.By using the skewness estimations of the cross-correlation data,we show that the simulated stochastic signals can be identified.Hopefully,it is useful for the detection of the stochastic HFGWs in future.