Zhongguo youzhi (Jan 2023)

植物油的红外光谱结合神经网络快速识别

  • 接昭玮, 刘卓，王继芬，古锟山，王之宇 JIE Zhaowei, LIU Zhuo, WANG Jifen, GU Kunshan, WANG Zhiyu

DOI
https://doi.org/10.19902/j.cnki.zgyz.1003-7969.210753
Journal volume & issue
Vol. 48, no. 1
pp. 79 – 83，93

Abstract

Read online

为实现对植物油的快速检测，借助衰减全反射-傅里叶变换红外光谱分析技术并结合深度学习算法对植物油开展光谱模式识别工作。实验获取8种植物油样本的光谱数据，采用标准正态变换和一阶导数预处理方法消除背景干扰，同时采用竞争性自适应重加权算法模型对各样本特征光谱数据进行提取，分别建立长短记忆神经网络（LSTM）、基于Levenberg-Marquardt算法改进的BP神经网络对提取特征波长后的植物油种类进行预测识别与比较，并采用后者进行了实际样品的识别检测。结果表明，通过提取特征波长，可有效提高LSTM模型的识别准确率，其最优准确率从提取特征波长前的30%~40%提高到80%~90%,模型运行时间从提取特征波长前的111 min 25 s缩短至 1 min 45 s。相较于LSTM模型,基于Levenberg-Marquardt算法改进的BP神经网络的分类识别准确率更高，达到99.852%，用于实际样品的识别，识别准确率达到100%。实验结果可为植物油的无损快速检验提供一定的参考与借鉴。 In order to realize the fast identification of vegetable oil, attenuated total reflection-Fourier transform infrared spectroscopy technology and depth learning algorithm were used to carry out spectral pattern recognition of vegetable oil. The spectral data of 8 kinds of vegetable oil samples were obtained in the experiment. The standard normal variate and the first derivative preprocessing method were used to eliminate the background interference. At the same time, competitive adaptive reweighted sampling model was constructed to extract the features spectral data of each sample. Long short-term memory neural network(LSTM) and BP neural network improved by Levenberg-Marquardt algorithm were established to predict and compare the types of vegetable oil after extracting feature wavelength respectively, and the latter was used for identification of actual samples. The results showed that the recognition accuracy of the LSTM model could be effectively improved by extracting the feature wavelengths, and the best recognition rate increased from 30%-40% before extracting the feature wavelengths to 80%-90%, and the model running time was shortened from 111 min 25 s before extracting the feature wavelengths to 1 min 45 s. Compared with the LSTM model, BP neural network improved by Levenberg-Marquardt algorithm had a higher classification recognition accuracy of 99.852%, and the recognition accuracy was 100% when used for the recognition of actural samples. The experimental results can provide some reference for nondestructive rapid testing of vegetable oils.

Keywords