Application of Transformer Model in Peak Correction of X-Ray Fluorescence Spectra

A novel estimation model based on deep learning is proposed for estimating the pulse amplitude of photons incident upon a silicon drift detector in X-ray spectroscopic systems. It is specifically designed for accurate analysis of spectroscopy where amplitude estimation of distorted pulses is an issue. A key step in the training of the estimation model is embedding of positional encoding and the calculation of multihead attention. This model uses the timing information in the input signal by adding positional encoding, improves the learning ability of the model, and realizes the accurate estimation of the amplitude parameters of the nuclear pulse signals. The simulation results show that the Transformer model can not only overcome the shortcomings of the traditional digital shaping methods in the inaccurate estimation of the amplitude of the distorted pulses but also improve the problem of the recurrent neural network (RNN) model’s poor ability to capture the remote information and can accurately estimate the amplitude parameters of the pulse with different degrees of distortion. During the experimental verification process, the peak area correction ratio is 93.01% in the region of interest (ROI) of Se, and the peak area correction ratio is 97.74% in the ROI of Sn. The model exhibits higher pulse estimation accuracy than other state-of-the-art pulse estimation methods and corrects the peak area in the ROI of characteristic peak while being much faster to compute.