Research On Automatic Measurement Method Of Quasar Spectral Lines Based On Machine Learnin

Quasars,one of the four major astronomical discoveries of the 1960 s,are important research subjects in the field of astronomy.They play a crucial role in the reionization history of the early universe,large-scale structures,intergalactic medium,and galaxy evolution processes.However,scientific questions concerning the formation of supermassive black holes in quasars and their interactions and evolutionary processes with host galaxies still require in-depth research.The measurement of quasar spectral features and the study of intrinsic physical parameters are crucial for addressing these questions and promoting our understanding of the universe.With astronomy entering the era of large-scale surveys and big data,projects like the Sloan Digital Sky Survey(SDSS)continuously release massive amounts of quasar data,providing unprecedented opportunities for relevant research.However,facing such vast data,the speed of scientific output gradually lags behind the data growth rate.Given the complexity of quasar spectral structures,efficiently and accurately measuring quasar spectral lines and obtaining their intrinsic physical parameters to accelerate scientific output has become a challenge for researchers.Although there are pipelines for automatically processing quasar data,there is still room for improvement.In recent years,artificial intelligence technology has been widely applied in astronomical research due to its powerful learning capabilities,but research on automatic quasar spectral line measurements is still insufficient.This thesis first measured and statistically analyzed the related physical parameters based on the latest SDSS quasar data,verified the reliability of the measurement results,and constructed a machine learning dataset with these results.To address the low accuracy of existing automatic data processing pipelines when dealing with complex quasar data,a U-net neural network was designed to identify anomalous data and absorption line data in quasar spectra,while a convolutional neural network was designed to automatically judge the quality of fit results.Based on these two models,an automatic quasar spectral line measurement pipeline was implemented.Additionally,a robust automatic photometry pipeline was designed and implemented to address issues encountered in long-term tracking observations of quasars and active galactic nuclei.The main work of this thesis includes:(1)A highly reliable quasar catalog was established as a sample set for subsequent research.Spectral features of 225,000 newly discovered quasars in SDSS-IV were measured,and the corresponding quasar catalog results were publicly released.Parameters of emission lines and continua such as C iv,Mg ii,Hβ,Hα,and [O III] were provided,and central black hole mass and Eddington ratio were estimated based on these parameters.Considering the complexity of quasar spectral radiation components,all fitting results were manually checked to ensure data accuracy.The obtained physical parameters were statistically studied,including the relationships between emission line velocity shifts,line widths,emission line luminosity,and continuum luminosity.The systematic differences between C iv and Mg ii black hole mass estimates were also revealed.The machine learning dataset in this thesis was constructed based on these measurement results.(2)A U-net neural network model was designed and implemented,effectively identifying anomalous data and absorption lines in quasar spectra.Using the 225,000 quasar spectral line measurement results as a dataset,a U-net neural network algorithm model was designed to remove anomalous data and absorption line data from the spectra,and to identify normal data for spectral fitting.Previous studies showed that more than 10% of the emission line spectral regions had strong absorption lines,severely interfering with the correct fitting of emission lines and the accurate acquisition of intrinsic physical parameters.Tests showed that the U-net model designed in this thesis achieved an accuracy of 99% for normal data recognition,with a false positive rate below 5%.The model effectively identified anomalous and normal data in the spectra,providing a reliable method for accurate automatic spectral line measurements.(3)A convolutional neural network model was designed and implemented,effectively judging erroneous fitting results during the automatic data processing.Using the225,000 quasar spectral line measurement results as a dataset,a convolutional neural network algorithm model was designed to automatically evaluate the quality of fitting results.Since emission lines require fitting multiple radiation components,improper parameter constraint ranges may lead to fitting failures.Previous studies showed that traditional fitting goodness indicators were inadequate for quality evaluation in large sample automatic data processing,but deep learning algorithms have advantages in such problems.Tests showed that the network model designed in this thesis achieved an accuracy of 97% for automatically judging erroneous fitting results,with a recall rate of 92%.Considering that the actual erroneous fitting results account for about10% of the total data,the model’s missed erroneous fitting results are estimated to be only 0.74% of the total data.The model’s performance basically met the requirements.(4)An automatic spectral data processing pipeline based on machine learning was implemented,effectively processing quasar spectra containing anomalous data and absorption lines.This pipeline was implemented based on the U-net and convolutional neural network models designed in this thesis.An automatic data processing test was conducted on 23,653 different emission line regions of 20,000 random spectra,showing that the pipeline could correctly fit 99.85% of the spectra.It effectively handled the complexity of quasar spectral structures,providing a guarantee for accurately obtaining quasar physical parameters.(5)A set of automatic photometry pipeline for image blur in tracking observations of quasars and active galactic nuclei with 1.26 m telescope is realized.This line is suitable for automatic metering of dim and weak signals and blurred images.After testing and analysis,it is found that the pipeline has a higher photometric success rate than the original pipeline of the telescope,and has the same accuracy as IRAF manual photometry,which provides convenience for the related scientific research of quasars and active galactic nuclei.In conclusion,based on the spectral line measurement of quasars,a reliable machine learning data set is constructed in this paper.At the same time,a large number of important physical parameters of newly discovered quasars are given,which provides more valuable data for the construction of larger unbiased samples and quasar-related studies.For the first time,U-net neural network and convolutional neural network are used to realize the pipeline of automatic measurement of quasar spectral lines,which significantly improves the efficiency and accuracy of data processing.A set of robust automatic photometry pipeline is developed to solve the problem of image blur in long-term tracking observation of quasars.The results of this thesis can provide more high-quality data in the field of quasar research,help to promote the in-depth development of related studies,and provide more powerful measured evidence support for revealing important issues such as the formation and evolution of quasars and supermassive black holes.