基于深度学习的 LSTM 模型在 X 荧光光谱中的应用 = Application of an LSTM model based on deep learning through X-ray fluorescence spectroscopy

• Main Authors: Tang, Lin, Li, Yong, Tang, Yufeng, Liu, Ze, Liu, Bingqi
• Other Authors: School of Electrical and Electronic Engineering
• Format: Article
• Language: Chinese
• Published: 2024
• Subjects: Engineering; Deep Learning; Long and Short Term Memory
• Online Access: https://hdl.handle.net/10356/173545
• Institution: Nanyang Technological University

在X荧光光谱分析中，为了解决传统谱分析方法中存在的特征峰计数率损失以及影子峰的问题，本文拟采用一种基于深度学习的长短期记忆（Long and Short Term Memory，LSTM）神经网络模型，该模型对核脉冲幅度时间序列具有较好的适用性，通过对样本的学习能够对核脉冲信号的幅度进行准确估计。鉴于核脉冲信号样本较大，模型训练效率低，特引入卷积神经网络（Convolutional Neural Network, CNN），利用其特有的卷积核结构逐层提取样本特征，能够有效减少样本数量，降低模型训练复杂度。使用粉末铁矿样品测量得到的一系列离线核脉冲序列产生模型训练所需的数据集，该数据集的64000个条目中，44800个用作训练集，12800个用作验证集，余下6400个用作测试集。实验结果表明：训练好的CNN-LSTM模型能够极大地节省训练时间，克服传统方法局部收敛的缺陷，也能够对不同程度畸变的输入脉冲进行准确的参数估计，在训练集和验证集上得到的准确率都高于99%。进一步分析计数修复结果，得到三个影子峰校正比例的平均值为91.52%，表明训练的CNN-LSTM模型对畸变脉冲产生的计数损失的校正比例约为91.52%。该模型能够有效校正因畸变脉冲幅度损失造成的影子峰，改善X射线荧光光谱中特征峰计数率精度，在X射线荧光光谱领域具有较高的应用价值。 [Background] Traditional X-ray fluorescence spectrum analysis has the limitations of poor accuracy of the characteristic peak counting rate and shadow peak. [Purpose] This study aims to propose a long and short term memory (LSTM) neural network model based on deep learning for the loss correction of the characteristic peak count rate and shadow peak. [Methods] Firstly, a LSTM neural network model based on deep learning was proposed to estimate accurately the amplitudes of nuclear pulse signals by learning samples. Then, a convolutional neural network (CNN) with unique convolutional kernel structure was introduced to deal with the challenges of large sample size of the nuclear pulse signal and the low training efficiency of the model by extracting the sample features layer by layer, thereby effectively reducing the number of samples and the complexity of model training. Finally, a series of offline nuclear pulse sequences of powdered iron ore samples were used to generate the dataset required for model training. Among the 64 000 entries in this dataset, 44 800 were used as training sets, 12 800 were used as validation sets, and the remaining 6 400 were used as testing sets. [Results] The trained CNN-LSTM model saves considerable training time, overcomes the defects of local convergence of traditional methods, and accurately estimates the parameters of input pulse under different degrees of distortion. Results show that the accuracy rate of the training and verification sets is greater than 99%. An analysis of the count repair results reveals that the average value of the correction ratio of the three shadow peaks, that is, the correction ratio of the depth learning model trained in this study to the count loss derived from the distorted pulses, is 91.52%. [Conclusions] The CNN-LSTM model can effectively correct the shadow peaks derived from the amplitude loss of distorted pulses and improve the accuracy of the characteristic peak count rate in X-ray fluorescence spectra. The model is shown to have high application value for the field of X-ray fluorescence spectroscopy.