Astronomical Transient Source Recognition Based On Deep Learning And Raspberry Pi

The astronomical transient source is the main research object of time domain astronomy,including the sudden emergence of natural targets such as supernovae and moving targets such as near-Earth objects and space debris.In the optical band,the large field of view telescope is the main observation device for the astronomical transient source,and has an unshakable position in the astronomical observation of time domain.At present,the main observation method of optical transientsource is still based on the combination of image subtraction,target automatic extraction and artificial recognition.Due to the need for manual intervention,this method of observation is inefficient,slow and accurate due to the experience of the observer.With the increase of data continuity requirements in astronomical observations in the time domain,large-field small-caliber telescopes distributed in different geographical locations have gradually become the main equipment for discovering and detecting astronomical sources.The automatic observation network composed of multiple large-field small-caliber telescopes can continuously search the sky to ensure the continuity and timeliness of observation.However,large-field small-caliber telescopes are scattered geographically,and the amount ofdata acquired is very large(hundreds of gigabytes to tens of terabytes/day).Geographical distribution is scattered.If data is transmitted back to the data center,network resources will be compared.High requirements,while reducing the timeliness of obtaining target information.In view of the need for real-time detection of transient sources in large-field small-caliber telescope networks,this paper has done the following work:First,on the original observation image,a large number of transient sources that can be used for classification are extracted by SExtractor source extraction algorithm.According to the symmetry of the optical system point spread function and the image signal-to-noise ratio,the target abnormal data is cleaned.Second,based on Convolutional Neural Network(CNN)and Recurrent Neural Network(RNN),structural improvement and a series of optimization training were carried out to generate two transientsource identification classifiers,and real data.More than 97% of the recognition accuracy is achieved.Third,using real data,it is found through experiments that CNN and RNN-based classifiers have different classification characteristics for different targets.The integrated learning method is used to fuse the results of the two classifiers to generate a new classifier,and the classifier accurate rate has increased to 99%.Fourth,under the Linux framework,the trained integrated classifier is loaded on the third generation Raspberry Pi through environment configuration.The Raspberry Pi can be directly bound to the telescope for real-time processing of data.The experimental test shows that the proposed method can realize the real-time classification of the transientsource on the terminal of the large-field small-caliber telescope,which provides a basis for the subsequent network-warning of the transient source observation network.