Neural networks for dimensionality reduction of fluorescence spectra and prediction of drinking water disinfection by-products
Neural networks for dimensionality reduction of fluorescence spectra and prediction of drinking water disinfection by-products
复制标题
DOI：
10.1016/j.watres.2018.02.052
复制
发表时间：
2018-06-01
期刊：
影响因子：
12.8
通讯作者：
Andrews, Robert C.
中科院分区：
文献类型：
--
作者：
Peleato, Nicolas M.;Legge, Raymond L.;Andrews, Robert C.
利用荧光数据结合神经网络提高饮用水消毒副产物（DBPs）的可预测性进行了研究。自动编码器处理高维荧光数据的新应用与并行因子分析（PARAFAC）和主成分分析（PCA）的常见降维技术有关。所提出的方法进行了评估的基础上组成的可解释性，以及预测有机物的反应性形成的消毒副产物。验证数据集上的最佳预测精度与自动编码器神经网络方法或通过利用全光谱而无需预处理。与其他方法相比，自动编码器的潜在表示似乎可以减轻过拟合。虽然DBP预测误差最小化的其他预处理技术，PARAFAC产生了可解释的组件，类似于从个别有机荧光团的荧光预期。通过分析网络权重，可以识别与DBP形成相关的荧光区域，这代表了区分荧光团分组之间的反应性的潜在方法。然而，不同的结果，由于应用降维方法进行了观察，指示需要考虑的作用，数据预处理的结果的可解释性。与目前用于DBP形成预测的常见有机措施相比，荧光显示出提高预测精度，当应用适当的预处理和回归技术时，DBP预测的改善最好地实现。这项研究的结果表明，神经网络的潜在应用，以最好地利用荧光EEM数据预测有机物的反应性的承诺。(C)2018爱思唯尔有限公司版权所有
The use of fluorescence data coupled with neural networks for improved predictability of drinking water disinfection by-products (DBPs) was investigated. Novel application of autoencoders to process high dimensional fluorescence data was related to common dimensionality reduction techniques of parallel factors analysis (PARAFAC) and principal component analysis (PCA). The proposed method was assessed based on component interpretability as well as for prediction of organic matter reactivity to formation of DBPs. Optimal prediction accuracies on a validation dataset were observed with an autoencoder-neural network approach or by utilizing the full spectrum without pre-processing. Latent representation by an autoencoder appeared to mitigate overfitting when compared to other methods. Although DBP prediction error was minimized by other pre-processing techniques, PARAFAC yielded interpretable components which resemble fluorescence expected from individual organic fluorophores. Through analysis of the network weights, fluorescence regions associated with DBP formation can be identified, representing a potential method to distinguish reactivity between fluorophore groupings. However, distinct results due to the applied dimensionality reduction approaches were observed, dictating a need for considering the role of data pre-processing in the interpretability of the results. In comparison to common organic measures currently used for DBP formation prediction, fluorescence was shown to improve prediction accuracies, with improvements to DBP prediction best realized when appropriate pre-processing and regression techniques were applied. The results of this study show promise for the potential application of neural networks to best utilize fluorescence EEM data for prediction of organic matter reactivity. (C) 2018 Elsevier Ltd. All rights reserved.