2019ApJ...873...17W

Traditional pulsar signal detection technology based on a fast Fourier transform (FFT) spectrum search and epoch folding requires a very long time to obtain an appropriate net signal-to-noise gain, especially for weak pulsar signals with low photon fluxes. We utilize a high-order spectrum with a nonuniform sampling strategy and an extra denoising process, including a high-pass filter and autocorrelation, to suppress the noise to a great extent. Because of the advantages of the deep convolutional neural network in two-dimensional data mining, the pulsar detection task is accurately realized, while the expert's subjective experience and the formal theory are avoided. The Rossi X-ray Timing Explorer data from three pulsars, PSR B0531+21, PSR B0540-69, and PSR B1509-58, are selected for the experiment, and the identification task is realized with a classification accuracy greater than 90%, with observation times of only 0.5 s, 40 s, and 15 s, respectively. Traditional methods have difficulty accomplishing the identification task within the same observation time. Further experiments reveal that the high-pass filter and autocorrelation can effectively suppress the cosmic background and random noise and that the nonuniform sampling of the bispectrum can avoid frequency leakage. Although the time complexity (O(N²)) of the algorithm is higher than those of the traditional FFT (O(N log N)) methods, the algorithm reduces the requirement of the observation duration time; thus, the computational complexity is comparable to that of traditional methods.