摘要
摘酒作为白酒酿造工艺中极其重要的一个环节,其实现方式通常采用人工判别。为实现白酒摘酒过程的智能化推进,提出一种傅里叶近红外光谱与二维卷积神经网络相结合的白酒基酒判别模型。模型原始光谱数据由傅里叶近红外光谱仪对采集的白酒基酒扫描所得,首先通过基于马氏距离的异常数据处理模块进行异常数据剔除,解决异常样本对于模型训练的影响,其次通过Kennard-Stone(KS)算法进行数据集划分,解决随机划分空间距离分布不均匀的问题,然后为满足二维卷积神经网络的输入条件,构建样本维度转换模块,之后通过图像数据增强模块解决样本数据较小的问题,最后再将处理后的图像数据集送入基于Resnet50的特征提取模型和基于Softmax的分类模型中进行训练。经预测,该模型对于白酒基酒四段判别的准确率高达97.58%,同时对比传统的决策树算法(TREE)、随机森林(RFC)、高斯朴素贝叶斯(GBN)、支持向量机(SVM)和多层感知机(MLP)等分类方法在各种不同条件下的最优结果,其准确率分别提高了6.46%、4.03%、25.81%、26.61%和23.39%。实验结果表明,傅里叶近红外光谱与二维卷积神经网络相结合可以有效地完成白酒基酒段数判别,为近红外光谱在智能化摘酒领域提供了一种理论可能。
Picking liquor is an extremely important part of liquor making process,which is usually realized by artificial discrimination.In order to realize the intelligent advancement of liquor picking process,a liquor base liquor discrimination model combining Fourier near-infrared spectroscopy and twodimensional convolutional neural network was proposed.The original spectral data of the model were obtained by Fourier near-infrared spectrometer scanning the collected liquor base liquor.Firstly,the abnormal data were removed by mahalanobis distance based abnormal data processing module to solve the influence of abnormal samples on model training.Secondly,the data set was divided by Kennard-Stone(KS)algorithm.To solve the problem of uneven distribution of random partition space distance,and then to meet the input conditions of two-dimensional convolutional neural network,build the sample dimension conversion module,and then solve the problem of small sample data through the image data enhancement module.Finally,the processed image data set was sent to Resnet50 based feature extraction model and Softmax based classification model for training.It is predicted that the accuracy rate of the model is as high as 97.58%for the four-stage discrimination of liquor base liquor.Meanwhile,compared with the optimal results of the traditional decision TREE algorithm(TREE),random forest(RFC),Gaussian Naive Bayes(GBN),support vector machine(SVM)and multi-layer perceptron(MLP)classification methods under various conditions.The accuracy rate increased by 6.46%,4.03%,25.81%,26.61%and 23.39%,respectively.The experimental results show that the combination of Fourier nearinfrared spectroscopy and two-dimensional convolutional neural network can effectively complete the discrimination of the number of base liquor segments,providing a theoretical possibility for the near infrared spectroscopy in the field of intelligent wine picking.
作者
翟双
张贵宇
庹先国
朱雪梅
罗林
ZHAI Shuang;ZHANG Guiyu;TUO Xianguo;ZHU Xuemei;LUO Lin(School of Automation&Information Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;Artificial Intelligence Key Laboratory of Sichuan Province,Sichuan University of Science&Engineering,Yibin 644000,China;School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China)
出处
《食品科技》
CAS
北大核心
2022年第9期250-256,共7页
Food Science and Technology
基金
四川省重大科技专项(2018GZDZX0045)
四川省科技成果转移转化示范项目(2020ZHCG0040)
四川省重点研发项目(22ZDYF0891)。
关键词
近红外光谱
白酒段数判别
二维卷积
near infrared spectroscopy
discrimination of liquor segment number
the two-dimensional convolution
