| 其他摘要 | The formation and evolution of contact binaries, one binary with a common envelope and strong material exchange and interaction between its two daughter stars, is still a subject of intensive study. With various survey projects (Kepler and TESS telescopes in space, ZTF telescopes on the ground), millions of light curves have been released, including hundreds of thousands of light curves for contact binaries. But how to identify, classify, and deriving the parameters of these huge amounts of data has clearly become a bottleneck problem. If the classical WD or Phoebe procedures is used for deriving the parameters of contact binary, the time of deriving parameters of one contact binary can take hours and days, which is clearly unaffordable. In this paper, the core of this thesis is the application of machine learning to the data processing of light curves of contact binaries, including identification, classification and deriving the parameters of contact binary. We have not only developed a complete machine learning-based method and process for processing the light curves of contact binaries, but also applied it to the TESS and ZTF observations to produce the corresponding parametric catalogues of contact binaries. We have also developed a method for the discrimination of member stars in open clusters and for deriving the parameters of contact binaries.The work in this paper includes the following.1. A method for identifying periodic variable stars based on LSSNR is proposed. The method is able to identify the periodic variable stars more accurately and give relatively accurate periods. After searching the light curve data of sectors 1-43 of the data released by the TESS telescope for periodic variable stars, a total of 26,206 periodic variable stars were obtained.2. A neural network-based classification model for periodic variable stars was developed. By Fourier transforming the light curves and selecting the amplitude and variable star period of the low-frequency terms as features, the data were downscaled and the neural network model was simplified, with the final F1-score average greater than 96%. Applying this method to the TESS survey data, the classification accuracy of 87.9% of these targets is greater than 90%.3. A fast batch of deriving the parameters of contact binaries method based on neural networks and MCMC is proposed and implemented. The regression neural network from the light curve to the corresponding parameter (inverse model) and from the parameter to the light curve (forward model) were trained using millions of theoretical light curves generated by Phoebe as training samples. The final parameters obtained for the contact binary include mass ratio, orbital inclination, temperature ratio, fill-out factor and third light ratio, as well as error ranges for the parameters. Using this method for deriving the parameters of contact binaries is four orders of magnitude more computationally efficient than before, allowing a target to be solved in twenty seconds, thus making it possible to derive the parameters of a large number of contact binaries. We have used this method to derail data released by the Kepler telescope and compared it with previous results, demonstrating that the method also performs very well in terms of accuracy.4. Based on our method, the TESS space telescope data and the ZTF ground-based telescope data are modelled separately, and their observations are identified, classified and deriving the parameters and finally the corresponding light curves are produced by the Phoebe program for comparison with the observations to verify the accuracy of the solution. The resulting parametric lists of TESS-joined binaries (699) and ZTF-joined binaries (86365) were produced. Finally, preliminary statistics on the parameter distributions are also presented and compared with previous findings on small samples.5. A DBSCAN-based method for the determination of open cluster members is developed, and a new method of deriving the parameters of contact binaries is applied to the cluster members. The method is validated with data from 30 open clusters, and the results show that the method is able to obtain more reliable cluster members, with detections up to about 21st magnitude. The identification of the contact binaries in the NGC 6791 cluster was performed, and the light curves of four EW-type stars were obtained, and the parameters of one contact binaries was derived by the method.In this paper, a number of machine learning methods (including unsupervised learning for dimensionality reduction and clustering, and supervised learning for classification and regression) are applied to the identification, classification and deriving the parameters of contact binaries, as well as to the identification of the cluster members, and a more complete set of methods for processing the light curves of massive contact binaries is proposed and some corresponding scientific results are obtained. These methods are also useful for the treatment of other types of binaries. |
修改评论