Transformer based on regression model for the denoising of X-ray pulsar profiles

To improve noise suppression and adaptability in pulsar observation profile denoising algorithms, this paper proposes a method utilizing a transformer based on regression model. Initially, pulsar photon time-series data are transformed into pulse profiles, followed by standardization and data augmentation. Subsequently, the Conv1d module extracts features directly from observed pulsar profiles and standard profiles, converting profile contextual information into intermediate semantic features. The Transformer module is then employed for deep feature extraction and regression analysis on profile features, resulting in denoised observed profiles. Crab and PSR B1509-58 data from the NICER telescope are used to validate the proposed method. The simulation experiments demonstrate that under the conditions where the observation duration of the Crab pulsar is less than 15 s and that of PSR B1509-58 is below 2500 s, the proposed method achieves average signal-to-noise ratios of 18.7796 dB and 10.1675 dB, average Pearson correlation coefficients of 0.9921 and 0.9482, and average phase offsets of 0.0417 ms and 0.6639 ms, respectively. These results outperform those obtained using wavelet transform denoising and kernel regression denoising methods.