Research On Modeling Method Of Near Infrared Spectroscopy Based On Deep Learning

Due to fast speed,high efficiency,simple operation,non-destructive and pollution-free,near infrared(NIR)spectroscopy analysis technology has been widely used in various industries such as food,fermentation,medicine and so on.An effective and accurate model is critical to NIR analysis technology.However,affected by measurement environment,experimental level,etc,traditional modeling methods perform not ideal because of insufficient feature learning ability and hard to traing.Deep learning algorithms have good deep feature representation capabilities,and they can automatically extract effective data features from a large number of inputs,so it is of significanc to apply the deep learning algorithm for data analysis of industrial processes.This dissertation researches the development of NIR model with deep learning algorithms.The research content includes the following aspects:(1)The near infrared spectral data contain a number of variables with large redundancy.In order to reduce the dimensionality of the spectral data while extracting deep features,Deep Belief Network(DBN)is introduced into NI modeling.This method first uses the Restricted Boltzmann Machine(RBM)to extract the deep feature of NIR data,and then employs the regression layer to take the place of the classification layer at the top of the network to achieve NIR modeling analysis.By applying this model to the detection of moisture content of pork minced meat and sugar content in citric acid wheat starch milk,and comparing it with traditional PLS method and BP algorithm,it can be seen that the prediction performace of DBN model and DBN-PLS model is obviously better than that of traditional modeling method.(2)Correlation between targeted values and spectral data is established by NIR model.However,due to the unsupervised pre-training mechanism of DBN,the learned features may contain a lot of information irrelevant to the targeted physical and chemical values.A deep belief network algorithm(VWDBN))based on variable weighting is proposed.By calculating the correlation coefficient between the physical and chemical value and the Restricted Boltzmann Machine(RBM)variables and designing the weighted reconstruction of the objective function,the model accuracy is guaranteed in extracting the characteristic variables related to the output.The improved VWDBN algorithm is applied to the citric acid wheat starch milk near infrared spectrum data,and the results show that the method effectively improves the prediction accuracy of the model.(3)Due to the aging of equipment and the change of environmental,the prediction accuracy of NIR detection is gradually deteriorating.The idea of real-time learning is introduced into DBN model to propose an update strategy of DBN model.The training sample data set with the maximum similarity to the current sample is searched in the historical sample database,and then the DBN is used to establish the model.At the same time,considering that the traditional selection of similar samples only considers the relation between the input data,the method of mutual information is introducd to select similar samples for the real-time learning algorithm.Firstly,the correlation between the measured spectrum and the historical sample base is calculated by using mutual information,and then the similar samples are calculated by the weighted degree of the spectrum.Finally,the DBN modeling algorithm based on mutual information is applied to citric acid wheat starch milk.