The calculation error of the stacked pulse amplitude generated by traditional pulse shaping methods leads to distortion in the X-ray fluorescence spectrum; thus, it is difficult to accurately analyze the spectrum measured in a high-stacking rate background.
This study aims to propose a transformer model based on deep learning for the pulse amplitude estimation of radiation measurements using high-performance silicon drift detectors.
Firstly, multi-head attention was applied to the transformer model, and an encoder-decoder structure with embedded positional encoding was employed to estimate the amplitude of stacked pulses. Then, a predefined mathematical model was used to simulate the pulse signal output by the detector for model training, and Gaussian noise corresponding to thermal noise and shot noise was added to the signal to simulate real nuclear pulses. Finally, experimental verifications were carried out on powdered iron ore samples and powdered rock samples, and relative error, corresponding to the accuracy of pulse amplitude estimation, was used as a model performance evaluation indicator.
Experimental verification results show that the average relative error obtained for eight offline pulse sequences of powdered iron ore samples and powdered rock samples is 0.89%, which means that the model can accurately estimate the amplitude of stacked pulses.