多环芳烃(PAHs)是煤, 石油, 木材, 烟草等燃料和有机高分子化合物等有机物不完全燃烧时产生的一种持久性有机污染物。 迄今已发现有200多种PAHs, 其中有多种PAHs具有致癌性。 PAHs广泛分布于我们生活的环境中, 水中的PAHs主要来源于生活污水, 工业排水和大气沉降。 使用三维荧光光谱法, 结合BP神经网络与交替三线性分解(ATLD)算法对水中的PAHs进行定性和定量分析。 以苊(ANA)和芴(FLU)2种PAHs为目标分析物, 用甲醇(光谱级)制备样本。 使用FS920稳态荧光光谱仪对样本进行检测, 设置激发波长为200～370 nm, 间隔10 nm记录一个数据; 发射波长为240～390 nm, 间隔2 nm记录一个数据。 设置初始发射波长总是滞后激发波长40 nm, 以消除一级瑞利散射的干扰。 随后使用BP神经网络法对待测样本数据进行预处理。 利用BP神经网络基于误差反向传播算法(error back propagation training, BP)原理, 对测得的三维荧光数据进行数据压缩处理, 该方法具有柔性的网络结构与很强的非线性映射能力, 网络的输入层、 隐含层和输出层的神经元个数可根据实际情况设定, 并且网络的结构不同时, 性能也有所差异。 随后, 用ATLD算法分解预处理后的三维荧光光谱数据。 采用核一致诊断法确定待测样本的组分数为2。 结果表明, ATLD算法分解得到两种PAHs(ANA和FLU)的激发、 发射光谱图与目标光谱非常相似, 能实现光谱重叠严重的PAHs(ANA和FLU)的快速定性和定量分析, 实现了以“数学分离”代替“化学分离”。 将预测样本导入训练好的BP神经网络中, 得到处理后待测样本数据的网络均方差(MSE)均小于0.003, 网络的峰值信噪比(PSNR)均大于120dB(数据压缩中典型的峰值信噪比值在30～40 dB之间, 越高越好), 可见BP神经网络对样本数据的压缩效果较好。 BP神经网络训练后, 得到输出值与目标值之间的拟合度高, 拟合系数达0.998, 具有较好的数据压缩效果。 使用ATLD算法对待测样本进行分解后得到平均回收率为97.1％和98.9％, 预测均方根误差为0.081 8和0.098 5 μg·L-1。 三维荧光光谱结合BP神经网络和ATLD能够实现痕量PAHs的快速检测。

Polycyclic aromatic hydrocarbons (PAHs) are persistent organic pollutants produced in case incomplete combustion of organic materials such as coal, petroleum, wood, tobacco, and other organic polymer compounds. More than 200 PAHs have been discovered to date, and many of them have carcinogenicity. PAHs are widely distributed in the environmentthat we live in. PAHs in water are mainly derived from domestic sewage, industrial drainage and atmospheric deposition. In this paper, three-dimensional fluorescence spectroscopy combined with BP (back propagation) neural network and alternating trilinear decomposition (ATLD) algorithm for qualitative and quantitative analysis of PAHs in water. In this paper, two PAHs, ANA and FLU, were used as analytes, and samples were prepared using methanol (spectral level). The samples were detected using a FS920 steady-state fluorescence spectrometer. The excitation wavelength was set at 200～370 nm, and data were recorded at intervals of 10 nm. The emission wavelength was 240~390 nm, and data were recorded at intervals of 2 nm. Setting the initial emission wavelength always lags the excitation wavelength by 40 nm to eliminate the interference of the first-order Rayleigh scattering. The sample data are then preprocessed using the BP neural network method. The BP neural network is used to compress the measured three-dimensional fluorescence data based on the principle of Error Back Propagation Training (BP). The method has flexible network structure and strong nonlinear mapping ability. The number of neurons in the input layer, the hidden layer, and the output layer can be set according to actual conditions, and the performance is also different when the structure of the network is different. Subsequently, the pre-processed three-dimensional fluorescence spectrum data were decomposed using the ATLD algorithm. Before the decomposition, the nuclear consistent diagnosis method is used to determine the number of components of the sample to be tested is 2. The results show that the excitation and emission spectra of ANA and FLU are very similar to the target spectrum, which can realize the rapid qualitative and quantitative analysis of PAHs (ANA and FLU) with severe spectral overlap. “Mathematical separation” replaces “chemical separation”. The predicted samples are imported into the trained BP neural network, and the network mean square error (MSE) of the sample data to be tested is less than 0.003, and the peak signal-to-noise ratio (PSNR) of the network is greater than 120 dB (typical peak signal in data compression). The noise ratio is between 30 and 40 dB, the higher the better. It can be seen that the BP neural network has better compression effect on the sample data. After BP neural network training, the fitting degree between the output value and the target value is high, and the fitting coefficient is 0.998, which has better data compression effect. Using the ATLD algorithm to decompose the samples to be tested, the average recoveries were 97.1% and 98.9%, and the predicted root mean square errors were 0.081 8 and 0.098 5 μg·L-1. Three-dimensional fluorescence spectroscopy combined with BP neural network and ATLD can achieve a rapid detection of trace amounts of PAHs.