Research On Pulsar Candidate Classification Based On Semi-supervised Learning

Identifying suspected pulsar signals from pulsar candidates is an indispensable part of the pulsar search process.The pulsar candidate identification methods based on machine learning developed in recent years is highly dependent on the marked samples,and still needs a lot of manual input in practical applications to obtain enough training samples;At the same time,the traditional pulsar candidate identification methods have the problem of model deviation due to the imbalance of positive and negative samples in the pulsar candidate data set itself.The semi-supervised generative adversarial network is able to effectively utilize a few labeled samples to achieve the identification research of unbalanced pulsar candidates,but this model still has certain shortcomings.Therefore,based on the semi-supervised generative adversarial network model,this paper has proposed two pulsar candidate identification methods based on semi-supervised learning to efficiently identify pulsar candidates.Pulsar candidate identification method based on semi-supervised attention generative adversarial network.The generator is constrained by the size of the receptive field,which makes it hard to learn the global features of the sample when the generative adversarial network composed of the convolution neural network is used to capture the sample characteristics.The discriminator is also limited to the receptive field,it can only predict samples based on local feature information,which makes some background information irrelevant to the classification task and interfere with the classification accuracy.For this kind of problem,this paper has integrated the attention module in the generative adversarial network,it can assist the generator in fully coordinating the correlation between the features of the convolution receptive field area and the remote area,so as to improve the generator’s ability to fit the real sample feature distribution;It has helped the discriminator learn important features related to the task and suppress irrelevant information through the attention weight coefficient,so as to optimize the screening ability of the model for pulsar candidates.Through the experiments,the identification effect of the proposed method has been improved in terms of recall rate and F1 score compared with the semi-supervised generative adversarial network model.Pulsar candidate identification method based on semi-supervised residual generative adversarial network.The semi-supervised generative adversarial network model has high requirements on the number of samples,especially the number of positive samples.When the number of labeled positive samples is insufficient,it will be difficult for the model to extract important features related to the task and affect the recognition effect.To solve this problem,based on the semi-supervised generative adversarial network,this paper has introduced the residual module in the generator to improve the ability of the network to capture the real sample distribution under the condition of small samples,so as to better train the classifier and better identify the pulsar candidates.The experimental findings indicate that the proposed method has further optimized its performance in terms of accuracy,precision,recall,and F1 score compared with the semi-supervised generative adversarial network model in the face of small sample data sets.