Variable Selection And Dimension Reduction For Soft Sensing In Fermentation Processes

In recent years,nondestructive testing technology,e.g.such as electronic nose and near infrared spectroscopy,has been widely used in detecting food,medicine and other processes for their non-destructive and mutual compatibility.However,electronic nose data and near infrared spectroscopy should be pretreated before modeling.Or,it may result in too many variables and high computational complexity.So,it is important to use dimension reduction algorithms to select appropriate variables for modeling.New dimension reduction methods of electronic nose data and near infrared spectroscopy are proposed,which focuses on solving problems of too many feature variables and high computational complexity.The main work in this thesis is as follows:(1)A two-level optimization method of sensor array is proposed to solve problems of cross sensitivity and information redundancy.In the first step,the principal component analysis and the similarity factor method are combined to optimize the sensor array,and then the sequential forward selection algorithm is used to optimize the sensor array.The method was applied to the monitoring of solid state fermentation of protein feed,Traditional feature reduction algorithms are also studied for comparisons.The three methods,artificial neural networks,support vector machine and linear discriminant analysis,were studied to compare prediction accuracy and the number of variables.The results show that proposed method has some advantages and applicability in optimizing sensor array.(2)In this work,a near-infrared wavelength feature selection method based on si PLS-LASSO is proposed to overcome the problem that spectral data has too many features and serious collinearity between spectral information.Partial least squares calibration model is established to predict pH values of the protein feed fermentation process.This method can overcome the shortcomings of LASSO and siPLS.Firstly,appropriate spectral intervals are selected by using siPLS;Secondly,LASSO is usedto select wavelengths in the selected intervals.Finally,the selected wavelengths are used for modeling.Case study shows the siPLS-LASSO method has obvious advantages in prediction and computation performance compared to traditional variable selection methods.So,siPLS-LASSO is an effective wavelength selection algorithm for modeling.(3)To extract deep information of near-infrared spectroscopy,this paper uses three typical models of deep learning,convolution neural network,deep Boltzmann machine and stack auto-coding network.To reduce the spectral data and get more concise and stable solid state fermentation prediction model,deep learning and genetic algorithm are combined to reduce the dimension of original near-infrared spectral and establish the support vector machines calibration model to predict pH values of the fermentation process.Case study shows that deep learning algorithm can automatically learn hierarchical feature representation and reduce the number of variables.At the same time,the combination method can help to further filter low-dimensional space variables,reduce time consuming and improve prediction accuracy.