Classifying Transit Light Curves Using Wavelet Analysis And Attention Convolutional Neural Network

With the development of astronomical observation equipment and technology,the large field of view telescope project has promoted the study of time-domain astronomy,among which the transit method for detecting exoplanets is one of the research hotspots in time-domain astronomy.The light curve is an important data that records the variation of the brightness of celestial objects over time,and contains rich physical information.With the improvement of the accuracy of time-domain astronomical data acquisition,the data scale has shown a significant increase,and traditional data analysis methods are no longer able to quickly and accurately search for target celestial objects in complex and diverse light curve sequences.In order to meet the processing needs of the new generation of time-domain astronomical observation data,efficient data processing and analysis methods are needed.This paper mainly focuses on the classification of light curve sequences in the transit method time-domain survey,and proposes corresponding solutions for the challenges faced by different data processing scenarios.The research focus and innovation of this paper include:(1)In the field of transit method-based time-domain surveys,to enhance network learning capabilities and handle complex and diverse light curve data with limited samples,a common practice is to employ phase folding to extract features,reinforcing periodic patterns and suppressing noise before network classification.However,the phase folding method fails to capture the multi-periodic and irregular variations of light curve sequences and relies on precise period parameters and extensive photometric data.To overcome these limitations,we propose a wavelet-based representation method for light curve time series.By utilizing a 6-layer bior3.9 wavelet with Sqtwolog hard threshold decomposition,this method effectively enhances periodic features,suppresses data noise,and reduces the original data dimensionality by approximately 32 times.Moreover,it separates high-frequency and low-frequency information by obtaining the decomposed components of the original light curve sequence.To address variable-length sequence inputs,we introduce cubic spline interpolation to resample the high and low-frequency components obtained through wavelet transformation,providing a novel approach for integrating neural networks with traditional features.(2)To handle the vast amount of data and complex nature of astronomical light curve sequences resulting from large field-of-view and high sampling rate time-domain surveys,we propose a convolutional neural network(CNN)classification algorithm incorporating temporal and channel attention.Leveraging the wavelet representation of light curve sequences,the high and low-frequency components of transit method-based time-domain survey data serve as dualchannel inputs to the network.By incorporating temporal and channel attention mechanisms,the network improves its ability to process long sequence data and capture the spatiotemporal correlations within light curve sequences.In testing using TESS photometric sequences,the network successfully identifies 94.32% of planets,90.71% of eclipsing binaries,81.10% of variable stars,and 95.16% of instrument noise and other categories,achieving an overall classification accuracy of 91.84%.This research provides effective solutions for processing and analyzing large-scale and high-dimensional light curve time series data obtained from transit method-based time-domain surveys,addressing the rapid and accurate identification of observed targets.Furthermore,it contributes novel insights and methodologies to the field of time-domain astronomy research.