应用FLS920P型荧光光谱仪对L-色氨酸溶液进行三维荧光光谱检测, 从中发现: L-色氨酸的特征荧光峰位于270/350 nm。 设定发射波长为350 nm, 测量激发谱。 由测量结果发现在250~260 nm区间, 谱线斜率较大、 线性度好。 因此选取250, 255和260 nm三个激发波长, 在每个激发波长下分别测量相应的荧光发射谱。 基于三条不同的荧光发射谱, 构建以激发波长为外扰变量的自相关光谱; 而以浓度为外扰变量的自相关光谱, 是以超纯水在不同激发波长的平均谱作为参考光谱, 通过参考光谱与样本平均谱的相关计算得到。 在此基础上, 将相关光谱数据分别与偏最小二乘回归(PLSR)和径向基神经网络(RBFNN)相结合, 建立溶液中L-色氨酸含量的预测模型, 研究结果表明: 采用浓度为外扰变量构造的荧光相关光谱信噪比较高, 建模的预测效果要好; 而在外扰变量相同时, 基于径向基神经网络建立的预测模型比基于偏最小二乘回归建立的预测模型对溶液中L-色氨酸浓度的预测结果更为准确。 其中, 以浓度为外扰变量时的径向基神经网络预测模型准确度最高, 该模型的预测相关系数为99.91%, 预测均方根误差为0.033 μg·mL^-1。 研究结果表明, 使用该方法能够对溶液中的物质含量进行准确测定, 可为食品安全监管提供帮助。

The three-dimensional fluorescence spectrum of the L-tryptophan solution was measured in a fluorescence spectrometer. The result showed that the fluorescence peak of L-tryptophan located at 270 nm/350 nm (excitation wavelength/emission wavelength). As the emission wavelength was fixed at 350 nm, an excitation spectrum can be measured. It can be found from the excitation spectrum that the curve has a high slope and good linearity in the range of 250~260 nm. Thus, the excitation wavelengths of 250, 255 and 260 nm were chosen to excite the sample, and three fluorescence emission spectra were measured. Based on the three spectra, the auto correlation spectra with disturbance variable of wavelength were constructed. In addition, the emission spectra of the ultrapure water under different excitation wavelength were measured and averaged to be reference spectrum. The auto-correlation spectra with disturbance variable of concentration were constructed by correlation calculation between the reference spectrum and the averaged spectrum of the samples. Combined the correlation spectral data with partial least squares regression (PLSR) and radial basis function neural network (RBFNN), the prediction models of the concentration of L-tryptophan were measured, respectively. The prediction results showed that the correlation spectrum constructed with disturbance variable of concentration had a better signal-to-noise ratio and better prediction performance. Furthermore, with the same disturbance variable, the model based on RBFNN was more precise than that based on PLSR. Among all the models, the model based on RBFNN with disturbance variable of concentration had the best prediction performance with correlation coefficient of 99.91% and root-mean-square error of 0.033 μg·mL^-1. This method can provide helps in the food safety supervision for the accurate determination of substances.