Pulsar candidate identification using advanced transformer-based models

Abstract

Listen

Translate article

Rapid and accurate identification of pulsars is a significant topic for large radio telescope surveys. With the enhancement of astronomical instruments, modern radio telescopes are witnessing an exponential increase in pulsar candidate detections. The application of artificial intelligence for the identification of pulsar candidates is an automated and highly effective solution to tackle the challenge of processing and recognizing vast volumes of data. In this work, using the data released by two surveys, the Commensal Radio Astronomy FasT Survey (CRAFTS) and High-Time Resolution Universe (HTRU), we propose a new framework to identify pulsar candidates. Firstly, due to the small number of real pulsars, we compare the performance of different data augmentation methods and find that the pulsar samples generated by the Deep Convolutional Generative Adversarial Network (DCGAN) based on deep learning techniques are closer to real pulsars. Secondly, we use two transformer-based classification models, Vision Transformer (ViT) and Convolutional Vision Transformer (CvT), to classify pulsar candidates, and find that the evaluation indexes of pulsar candidate classification based on two transformers can reach 100%. Finally, we use the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm to visualize the results of our identification framework. The results showed that pulsar and non-pulsar samples are separated from each other in multidimensional space. Therefore, it is a new attempt to apply transformer technology to pulsar candidate classification, and it could be of great significance to subsequent theoretical research.