Generally, pulse truncation events caused by measurement systems often present challenges to pulse height analysis in the field of spectroscopy and radiometry, resulting in spectral distortion.

This study aims to propose a composite neural network model for accurately estimating the heights of truncated pulses.

Firstly, a long and short-term memory (LSTM) network was embedded into the UNet structure to construct a composite neural network model (LSTM-UNet). Then, the model was trained for height estimation of truncated pulses output by silicon drift detectors using a simulated pulse dataset for which the pulse amplitude matrix superimposed with noise was taken as input signal while the output signal was a set of expanded pulse heights. Finally, the performance of the model using relative error indicators was evaluated by analyses of powder iron ore and powder rock samples.

The average relative error of the UNet-LSTM model for pulse height estimation analysis on simulated pulse sequences is approximately 2.31%, which is 1.91% lower than the average relative error of traditional trapezoidal shaping algorithms. Verification results of the UNet-LSTM model on measured pulse sequences with different degrees of truncation show that the average relative error obtained during the height estimation of two samples and eight sets of offline pulse sequences is 2.36%.

The results reveal that the proposed model can accurately estimate truncated pulse heights.