基于连续投影算法的食用油激光诱导荧光光谱特征波长筛选

针对目前质量监督领域中难以快速准确地识别食用油种类的问题,提出了一种激光诱导荧光技术结合连续投影算法的食用油光谱识别方法.实验搭建激光诱导荧光系统采集了5种食用植物油共计500组荧光光谱数据.首先,通过实验对比选择Savitzky-Golay卷积平滑算法对荧光光谱进行预处理;然后使用连续投影算法筛选特征波长;最后,将筛选出的特征波长送入建立的概率神经网络模型中进行训练和测试,并通过不同时间采集的油样组成独立验证集进行再次验证.结果表明:通过特征波长筛选,从全光谱2 048个波长中优选出11个,减少了冗余信息,波长数缩减为原来的0.54%.并且在概率神经网络模型中的训练样本准确率和测试样本准确率分别达到100%和95%,效果好于径向基函数神经网络和BP(Back Propagation)神经网络.在独立验证集的预测准确率也达到了91%.因此,将连续投影算法用于食用油激光诱导荧光光谱特征波长筛选并结合概率神经网络模型可以实现食用油的快速准确分类且具备通用性,并为进一步设计专用的在线式食用油种类识别仪提供了理论依据.

Continuous projection algorithm for screening characteristic wavelength of laser induced fluorescence spectrum of edible oil

Aiming at the difficulity to quickly and accurately identify the type of edible oil in the current quality supervision field, a laser-induced fluorescence technology combined with continuous projection algorithm is proposed to identify the edible oil spectrum. In the experiment, a laser-induced fluorescence system is set up to collect a total of 500 sets of fluorescence spectrum data of 5 edible vegetable oils. First, the Savitzky-Golay convolution smoothing algorithm is selected through experimental comparison to preprocess the fluorescence spectrum; then the continuous projection algorithm is used to filter the characteristic wavelengths; finally, the selected characteristic wavelengths are sent to the established probabilistic neural network model for training and testing, re-verifying through the independent verification set of oil samples collected at different times. The results show that through the screening of characteristic wavelengths, 11 are selected from the 2 048 wavelengths in the full spectrum to reduce redundant information and the number of wavelengths is reduced to 0.54% of the original. And the accuracy of training samples and test samples in the probabilistic neural network model reached 100% and 95%, respectively, which are better than radial basis function neural networks and Back Propagation neural networks. The prediction accuracy rate in the independent validation set also reached 91%. Therefore, using the continuous projection algorithm to screen the characteristic wavelengths of the edible oil laser-induced fluorescence spectrum combined with the probabilistic neural network model can realize the rapid and accurate classification of edible oils and is universal, and provide further design of a dedicated online edible oil type identification instrument The theoretical basis.