Research On Pulsar Candidate Selection Method Based On Artificial Intelligence

Pulsar candidate selection is an important step in the search task of pulsars.With the increasing scale of pulsar surveys and increasing sensitivity of modern radio telescopes,more and more pulsar candidates have been produced,but there is a lot of radio frequency interference or background noise,some of which are very similar to pulsar signals.The screening method of manual recognition is very inefficient and subjective,which cannot meet the time-sensitive requirements of survey data,and cannot realize real-time processing of data,so it is difficult for manual recognition to select candidate which is worth further observation.Therefore,it is very important to search new pulsar by constructing an intelligent and efficient pulsar candidate selection model which can improve candidate selection performance.Artificial intelligence is a method for real-time,efficient and accurate screening of pulsar candidates.As a classical algorithm in artificial intelligence,neural network has achieved good results in pulsar candidate selection.However,the severe class imbalance in pulsar candidates will cause neural network to tend to the prediction of majority class(non-pulsar signal)in training,which makes the recognition rate of minority class(pulsar signal)low.In addition,feature selection and stability of neural network also need to be considered when building the screening model.To select real pulsar candidates from the massive data as much as possible,improve the accuracy and reduce the workload of further verification,this thesis conducts research on the selection of pulsar candidates based on artificial intelligence method,as follows:(1)A pulsar candidate selection method based on self-normalizing neural networks is proposed.This method uses self-normalizing neural networks,genetic algorithm and synthetic minority over-sampling technique to improve the screening performance of pulsar candidate.The self-normalizing property of self-normalizing neural networks overcomes the vanishing and exploding gradients in training process of deep neural networks,which greatly accelerates training process.To eliminate redundancy of sample data,genetic algorithm is used to select the sample features of pulsar candidate to obtain optimal feature subset.Concerning the severe class imbalance caused by limited number of real pulsar samples in data,synthetic minority over-sampling technique is used to generate pulsar candidate samples,which reduces the class imbalance rate.The experimental results on three pulsar candidate datasets based on classification accuracy show that the proposed method can effectively improve the screening performance of pulsar candidate.(2)A pulsar candidate selection method based on conditional least squares generative adversarial network is proposed.This method uses a framework that combines conditional adversarial networks and least squares adversarial networks to enhance the ability to screen pulsar candidates.Concerning the severe class imbalance caused by limited number of real pulsar samples in data,the generative adversarial network is used as a generation model to reduce imbalance rate.The least square loss function is used to improve the stability in training process of network,and minority class are generated by adding conditional data.The convolutional neural network is used as classifier of pulsar candidate,which can automatically to learn discriminant characteristic of the pulsar candidate and output result of selection.Experimental results on HTRU medlat dataset show that the proposed method is a deep learning method which can efficiently and accurately screen pulsar candidates.