Accurate identification of X-ray single photons using multi-dimensional features and positive-unlabeled learning

Abstract

Traditional thresholding for X-ray astronomical detection faces challenges in balancing low-energy photon detection efficiency and noise suppression. This paper proposes an innovative algorithm based on multidimensional feature engineering and positive-unlabeled (PU) learning. The algorithm constructs morphological features of photon events, greatly enhancing the ability to distinguish photons from noise. By combining the PU learning framework with an adaptive threshold, the algorithm addresses the identification of low-energy photons. Experimental results show that its ability to distinguish photons from noise (d’ value of 1489.49) far surpasses the upper limit of conventional methods (maximum d’ value of 11.13). The algorithm accurately recovers the low-energy spectrum, which traditional methods severely distort. This leads to more than an order of magnitude improvement in spectrum recovery accuracy (NMSE). The algorithm demonstrates excellent performance in single-photon identification, providing key technology for realizing the potential of CMOS detectors in space exploration.