• Application of a neural network model with multimodel fusion for fluorescence spectroscopy

    分类: 核科学技术 >> 辐射物理与技术 提交时间: 2024-06-19

    摘要: In energy-dispersive X-ray fluorescence spectroscopy, the estimation of the pulse amplitude determines the accuracy of the spectrum measurement. The error generated by the amplitude estimation of the pulse output distorted by the measurement system leads to false peaks in the measured spectrum. To eliminate these false peaks and achieve an accurate estimation of the distorted pulse amplitude, a composite neural network model is proposed, which embeds long and short-term memory (LSTM) into the UNet structure. The UNet network realizes the fusion of pulse sequence features and the LSTM model realizes pulse amplitude estimation. The model is trained using simulated pulse datasets with different amplitudes and distortion times. For the pulse height estimation, the average relative error of the trained model on the test set was approximately 0.64%, which is 27.37% lower than that of the traditional trapezoidal shaping algorithm. Offline processing of a standard iron source further validated the pulse height estimation performance of the UNet-LSTM model. After estimating the amplitude of the distorted pulses using the model, the false-peak area was reduced by approximately 91% over the full spectrum and was corrected to the characteristic peak region of interest (ROI). The corrected peak area accounted for approximately 1.32% of the characteristic peak ROI area. The results indicate that the model can accurately estimate the height of distorted pulses and has substantial corrective effects on false peaks.