A Pulsar Search Method Combining a New Feature Representation and Convolutional Neural Network*

Abstract

Abstract The radiation energy of X-ray pulsars is mainly concentrated in the high-energy ray bands, so processing high-energy photon signals is helpful for discovering some young and active pulsars. To quickly and accurately detect effective pulsar signals from a large number of samples within a finite observation time, an automatic identification algorithm for pulsar candidates based on X-ray observations is developed in this paper. First, the autocorrelation operation is used to improve the signal-to-noise ratio of the profile and solve the initial phase misalignment problem. Then, the candidate frequency range is expanded, and the output signal is folded according to these frequencies to obtain a series of profiles. The six statistical features of these profiles are extracted to generate frequency-feature curves. Compared with the traditional epoch folding method, the frequency-feature curves show more consistent characteristics. To improve the classification accuracy, the frequency-feature curves are converted into two-dimensional images, and ConvNets are used for deep feature extraction and classification. A simulation method based on the nonhomogeneous Poisson process is utilized to create the training set, and generative adversarial networks are used for data augmentation to solve the class imbalance problem caused by limited pulsar samples. Finally, the RXTE observation data of PSR B0531+21, PSR B0540-69, and PSR B1509-58 are selected for testing. The experimental results show that the highest recall and precision reached 0.996 and 0.983, respectively. Demonstrating the considerable potential of this method for identifying pulsar candidates based on X-ray observations.